Blog

Head of Post Production – Pedestrian.TV

Melbourne

Posted 14 Sep 2022
Closes 14 Oct 2022
Job description ┃Head of Post Production (Technical finisher)
Employment:
Part-Time Hybrid: 3-4 days per week. 
Location:
This role will be a hybrid local x remote with our office in Brunswick East, Melbourne.
We prefer to have someone based in Melbourne to help with computer and server management, however, our industry leading server ecosystem enables all of us to work remotely, so interstate unicorns will be considered. 
About Enamoured Iris
E.I. is a company that specialises in communicating with Millennial & Gen Z customers; by fusing the art of cinema and the science of marketing, we’re inspired to create unskippable content to help audiences remember brands. People describe us as being bold, provocative, playful, real, and our content as being adventurous, rebellious, creative, and innovative.
 
 
OUR VALUES
Mastery Stop at nothing to deliver your best.
Playfulness Only take your work seriously, and never yourself.
Discovery Always be open to exploration and never bound by convention.
Collaboration Remember teamwork makes the dream work.
Imagination Think big, pull back later.
 
 
Type of people who work here: 
Passionate, driven, autonomous, unique, future-focused and self-aware.
United by a love of film and advertising, our team is highly collaborative and inspired to create cinematic video marketing campaigns that leverage off modern advertising tactics. 
WHAT WE ARE LOOKING FOR
You’re a finisher, you see the missing pieces of a project and can take control to deliver to an excellent standard. You have an eye for detail, and are over 50% technically minded. 
This role will play an imperative part in maintaining Enamoured Iris’s high-quality production benchmarks within the travel, music and lifestyle industries, and will ensure the right production resources are in place to achieve business goals and quality creative outcomes for clients.
If you’re applying, you’re not just a good editor, you’re a great editor. You’re equally passionate about sound, visuals, and narrative. When we say passionate, we’re not saying you really enjoy balancing these three elements, we’re talking about ‘OCD passion’, that triggers a shiver down your spine when an assistant editor doesn’t cross-fade your sfx clips, and leaves the audio clipping in the red-zone. 
You’re the type of person who wants to learn more about new technologies and loves to experiment with videos to make them unique. 
Your saved Instagram posts have a healthy pool of emerging artists, inspirational content that you are willing to share with the team. 
Furthermore, you have a general interest in the art of special effects and motion graphics. You possess the basic knowledge of various post-production software to collaborate with with 3d and motion graphic artists.
If you have studied an instrument in your lifetime, please highlight this in your application, as an understanding of music really positions you at the front of the queue. 

Your Personal Characteristics
Passion for traditional film techniques, excellent attention to detail and process orientated, highly practical, clear communicator, excellent organisational skills, team player and, logical.

Job Overview
The Head of Post Production will be a master of technical finishing, responsible for the day to day and overall strategic development of the Post Production department. 
This role will be both managerial and creative, you will be assisting the producers to prepare your team for the weeks ahead and you will be there, boots on the ground, ensuring the final product is finished to an immaculate standard. 
The primary function of this role is to ensure that the department stays within budget and on schedule while maintaining the highest technical and creative standards. This role leads, supports and co-ordinates the editing team and ensures technical operations and workflows are well maintained. The Head of Post Production will work closely with the managing director to implement new technologies across the department when required. 
Daily Life (Responsibilities and Duties)
Qualifications
Sign up with your Facebook
or Linkedin account
Please select at least one of the following options to continue
By signing up, you agree to Pedestrian Group’s Terms of Service and consent to our Privacy Policy.
Already registered? Login here

source

‘Elvis’ Music Producer Dave Cobb on Recreating Classic Songs: ‘It’s Terrifying to Even Attempt to Do That’ – Variety

Variety Print Plus Subscriber
Variety Print Plus Subscriber
Variety Print Plus Subscriber
Variety Print Plus Subscriber
Variety Print Plus Subscriber

By Chris Willman
Senior Music Writer and Chief Music Critic
Renowned music producer Dave Cobb was working on a film job where he recreated the music of one singing icon — Tammy Faye Bakker — when he got the assignment to do the same for an even more significant vocalist, high a bar as that may have been: Elvis Presley. Truly, both figured into his youth, growing up as a Pentecostal kid in the Southeast. But even if one were to consider these singers on the same plane, “Elvis” would require quite a lot more out of him than “The Eyes of Tammy Faye,” as workloads go, due to the sheer amount of Presley catalog being recreated. Ultimately, he wound up working on 80 different tracks for the Baz Luhrmann film.
He admits he was intimidated by the task, although probably not as much as a producer whose parents hadn’t constantly played Elvis records from the time of his birth might have been. It also doesn’t hurt, perhaps, that Cobb has worked as the constant collaborator of a lot of modern greats, including Brandi Carlle, Jason Isbell and Chris Stapleton, as well as having started moving into film music more, starting with 2018’s “A Star Is Born,” for which he produced Lady Gaga.

The later parts of the film use actual vintage Presley recordings, but it was Cobb’s role to produce and wholly recreate the “young Elvis” tracks sung by Austin Butler, as well as come up with the music representing Presley’s many Black influences in the film, and some vintage country as well — and then do one wholly modern-sounding number with Stevie Nicks and Chris Isaak. How’d he do? The proof is in the pudding that most moviegoers don’t even notice the transitions between the newly recreated Butler tracks and the original masters heard toward the end of the film. He came, he conquered, he played house.
Variety caught up with Cobb just after the film opened to considerable enthusiasm from audiences.
There is a list of 80 tracks you worked on for the movie, either producing or mixing. Not every one of those is in the film or on the soundtrack; some of those are alternate arrangements and some are things that didn’t make the final cut. Still… Not many people would take on doing 80 songs for one movie.
No. [Laughs.] But it wasn’t all in one fell swoop, thank God. We started this pre-pandemic, and then everything got shut down for quite a while. We picked back up about a year ago. So 80 credits was easy to manage over three years, you know? It was pretty incredible, because some of those days at RCA, we had 15 people coming to sing different characters. We’d have Les Greene coming in to do Little Richard and Yola to do Sister Rosetta Tharpe — just one after the other of incredibly talented people who got behind a microphone and made this happen. The band’s playing completely live. Every bit of the music that I worked on in the film is completely live, with everything happening at one moment.
We had to find a Little Richard, a Rosetta Tharpe, an Arthur Crudup, and that was part of the gig as well, and the fun, trying to find people who could pull it off. Baz found Austin, and God, what a find that was — Austin really sells Elvis vocally. But trying to find Little Richard is one of the most difficult things in the world; he’s got one of the most unique voices of all time, so that was extremely difficult. Sister Rosetta Tharpe, extremely difficult, so I thought of Yola, because Yola can do anything. She walked in, dressed for the part with the guitar and everything, and absolutely overdelivered. And the same goes for finding all the gospel stuff in there. Baz from early meetings really wanted to show where Elvis got everything from, and bringing people like Shannon Sanders in – he’s a vocal arranger and choir director in Nashville — was huge as well… And the Settles Connection and the McCrary sisters, as well as Robert Randolph and his Family Band, all these incredibly talented people making these scenes happen with all the gospel stuff as well.

A lot of that casting was pure vocal casting, where the actor would be lip-synching to a more trained singer’s vocal. But then there were a couple of instances where the vocal casting became the on-screen casting, as well.
That’s correct. Yola, who I’d gotten to work with her briefly on the Highwomen, came in dressed as Rosetta Tharpe with a guitar and the whole nine, and just obviously had the voice, and she overdelivered. Baz fell in love with her immediately and wound up casting her as the part in the film. And Shonka Dukureh was recommended by Odesa Settles from the Settles Connection to come and be part of the choir. She started singing as we were doing these revival scenes and everybody’s like, “Oh my God.” So eventually she got cast to be in the film as well, as Big Mama Thornton. Shannon Sanders, the choir director, played such a big part that Baz cast him as well — he’s the Pentecostal minister in the tent scene.
How much did you get done before the big pandemic interruption?
Quite a bit. We had done quite a bit of the backing tracks for the ‘50s and ‘60s Elvis stuff, and recorded just a myriad of people, including Yola, in that batch. And then we actually had the idea to make it really authentic when we recorded the tent revival scene. We wound up driving about an hour outside of Nashville to this old, little 1800s Pentecostal church, and actually recorded that (audio) in the church. It was supposed to be reenacting what a tent revival would be like, but it actually turned into a real service. Austin and Baz and Elliot Wheeler (the executive music producer) came down and the whole team was there. The switch (was flipped) from just playing like you were in church to church. The band just wouldn’t stop playing, and it was beautiful. We captured so much gospel content right then and there. It just kept going and going and going and it was a real revival. We did, God knows how many songs in that, because it just kept one after the other rolling into each other.
Speaking of gospel, the last time we talked to you was about your work on the “Eyes of Tammy Faye” movie.
That’s right. It seems to be a theme, right? “Eyes of Tammy Faye” was right before we started on “Elvis.” I grew up Pentecostal, so I have such a love for gospel music, and it’s great to see two extremes of it, between “Tammy Faye” and this film.
People may be surprised at some of the more obscure songs that are in the film or on the soundtrack, like one I can say I didn’t know, “Cotton Candy Land.”
Honestly, I didn’t know that either. Baz and his team found that, and it’s such a really obscure, deep-dive Elvis track. That was cut a couple of times. It’s in the score with Stevie Nicks singing. And then later on, Baz really wanted a remix, so we did one where we have Chris Isaak and Stevie Nicks singing this kind of dance version of the track. Actually, it’s more like Serge Gainsbourg meets Elvis meets Nancy Sinatra or something.
Stevie is (vocally) playing a character in that too, by the way, which is really cool. She got really into character, because at the very opening of the film (where her voice appears on the soundtrack), you see a gypsy. So Stevie’s singing as a character, which is really interesting.
And doesn’t Chris Isaak also appear on the soundtrack as the voice of Hank Snow?
Yes. He’s singing the song “A Fool Such as I” in the film (for the Snow character, portrayed by another actor). It was funny because as we were working on the film, Elliott Wheeler calls me and is like, “Do you know anybody who could do Hank Snow?” And I was in the studio with Chris doing “Cotton Candy Land,” and Chris says, “I can do that.” Chris knew the song that they wanted and literally immediately started singing it on the spot, and just nailed his voice. I mean, it’s unbelievable. Chris can do anything. He’s one of my favorite people in the world.
Luhrmann clearly had a lot of strong ideas about the music of the film and wasn’t just leaving you guys to your own devices to develop it.
He is a genius. I mean, I’ve never met somebody who, when they talk about things, you visually see what they’re saying all the time. When he was in the studio, as we were doing the tracks, he was directing the scene in front of our eyes, and I think everyone played better because he was there. I think everybody understood the intent behind every note they played because he was there. He’s such an incredibly talented director, but the guy could be the best music producer in the world too, because he connects with people like no other. I was in awe of him being in the room and really getting the best of it out of everyone.
Can you talk about why Austin worked out so well?
Austin, the first time I met him, he walked in, and we had this huge studio band ready to track this stuff, and he gets behind the mic and starts singing and was just so impressive. I would’ve been scared to death, walking in a large studio with a bunch of session players and Baz and Elliot Wheeler and Jamieson Shaw and all those guys. To a certain extent when he came in the studio for the first time, he was talking like Elvis, and obviously had really researched, researched, researched and had an incredible grasp on Elvis even at that point —before they’d started filming. He really killed himself to get Elvis and nail Elvis’ spirit and intent and his emotion and his personality — and the vocals. I saw him at the Memphis premiere a couple weekends ago and the first thing I said was about how I was mixing the soundtrack and I just couldn’t believe I wasn’t pulling up Elvis’ vocal.

Were you nervous at all about remaking the classic tracks and not getting it quite right?
Anytime you think of trying to recreate such iconic recording, it’s absolutely terrifying to even attempt to do that. So, yeah, I was super scared about it all. But the musicians all came in really prepared. And we actually mic-ed everything and used equipment that would’ve been used in that era. We did it to analog tape, and we had ‘50s mic preamps and ‘50s ribbon microphones. The Country Music Hall of Fame has some of the equipment that Elvis used, including the slap you hear on some of those early records, and they were kind and let us borrow that. So we incorporated as much authenticity as we could into the recording side, even aside from the musicianship and the playing. So much attention to detail went into not faking it, but actually doing it the same way they would’ve done it. But just because you have all those ingredients to make it happen doesn’t mean you’re gonna nail it. And everybody poured their heart in to kind of make it come to life and actually sound like those records as much as we could.
In mixing something like this, it seems like there’s probably a fine balance to strike between wanting authentic period sounds and then gearing it toward however many speakers the average movie theater has right now — it’s going to be disappointing to the mass audience if you have these big, explosive scenes and it truly sounds just like a Sun record playing on a 45 spindle.
Stuff that I mixed, we really had really a pretty accurate balance and picture of what it was gonna sound like as it was tracking. When they were making Sun records, I’m sure they were going early on directly to mono and mixing on the fly. So I had an incredible team of engineers, Darrell Thorpe and Greg Gordon, mixing as as it was going down. When a solo happened, they were riding the mic preamp. And the players also paid homage to how players played back then. When they needed to get loud to get loud, they’d get loud; if they needed to get quiet, they’d get quiet. So a lot of the mixing happened completely live. And when I got the files, really the only things I did to make it bigger was maybe in the bottom end, to add kind of sub to a kick drum, which you don’t hear on a lot of old records; you don’t really hear the kick drum. So there are things like that that we modernized, but overall, it’s as authentic as possible.
Some of the stuff that we we talked about, especially when you see “Baby, Let’s Play House” in the film, and they’re playing live, and it’s really hard to get good recordings or hi-fi recordings of Elvis playing live in the ‘50s. And I think that was fun, to kind of capture the energy that might have been, but with a bit more fidelity. That was really striking a balance, in having that thing where everything sounds like it’s blowing up, like a great ‘50s record, but also it’s a little more expansive and wider than the original records. Which I think are perfect. We probably shouldn’t even try. But we did slide that a little bit.
What do you most appreciate about the way music is represented in the film?
You know, Baz had the whole idea musically for how everything fits together and and the choice of songs and the choice of legends that are in it. Seeing the fabric of American music all come together in a few days for the majority of the soundtrack was pretty magical. You see the influence of the church. You see the influence of blues, you see the influence of soul. You see the crooner side, and you see the birth of rock ‘n’ roll almost recreated in front of your eyes in the studio. Some of those beats and shuffles and things like that are long forgotten, and no one really uses that, but my God, how powerful are they? There’s a reason why Little Richard has so many hits; it just feels great. There’s a reason why Elvis has so many hits. To see that music happening organically and live with real human beings playing it and going for it, I think that’s the biggest takeaway. There’s definitely not a computer making it happen.
We should probably ask you who some of the musicians are. Do these arrangements just automatically come to them— “Oh yeah, this is how that Little Richard shuffle goes” — or is there a level of scientific analysis to try to reverse-engineer how they came up with some of those parts?

I picked people who I knew would get it, who had all that stuff in their formative years of playing. They even hyper-criticized the parts just to make sure it was as close as possible. One of the players I had to give a lot of credit to is Chris Powell, the drummer. I mean, he was playing on Little Richard, he’s playing on Rosetta Tharpe, he’s playing on the Elvis stuff, nd he was able to switch up and nail the variety of music that we had to do. As well, this guy JD Simo, who’s playing electric guitar — he’s playing the Rosetta Tharpe guitar parts, he’s playing Scotty Moore, he’s playing James Burton. It’s just incredible how they were able to be so versatile on the things we did on the spot.
Even though we had ‘50s Elvis period tube equipment and microphones, when we went to ‘60s Elvis, we were changing it all to solid-state transistor stuff and solid-state microphones, and just trying to really chase the audio archeology of these records. And I think we did our damnedest just to get that right.
Was Elvis part of your upbringing and youth?
He’s huge, huge to me. You know I grew up Pentecostal, and we had to go to church it felt like every single day. And we had a lot of (gospel) quartet records — and Elvis did quite a bit of gospel stuff too. But for some reason, (secular) Elvis and Buddy Holly snuck through the gates of what we listened to, and Elvis was my dad’s favorite. So there was always Elvis going, all the time. My mom told me that when my dad was young, he tried to impersonate Elvis vocally all the time. So I always heard Elvis, always, as a gigantic part of growing up. I may not have heard Black Sabbath or Led Zeppelin early on, but I heard Elvis since birth.
And I think Elvis sums up the Southeast region. I think most people who grew up in the Southeast have the same kind of church overtones in that kind of mix of soul music, country music and gospel music; it just seems to swim around here. So growing up in the Southeast, I definitely felt a kinship to Elvis, particularly the gospel stuff too. So this is a gig I never thought I would get, and it was a dream to work on.
Anything else in the hopper, film-wise, for you yet?
I don’t know if I’m supposed to talk about it yet. But I’ve really tried to make a conscious effort to really go into film; in the last three years, after “A Star Is Born,” I’ve just really gone pretty, pretty deep into it. I’m fascinated with it and I love it. It’s a different challenge and a lot of fun to get to do such variety when you do film. Sorry, no pun intended.
A Variety and iHeartRadio Podcast
The Business of Entertainment

source

Vectorworks, Inc. Launches 2023 Version Of BIM And CAD Product Line – Live Design

source

Avid Learning Academy Program Upgrades to Pro Tools Artist – StreetInsider.com

FREE Breaking News Alerts from StreetInsider.com!
StreetInsider.com Top Tickers, 9/15/2022
BURLINGTON, Mass., Sept. 14, 2022 (GLOBE NEWSWIRE) — Avid® (NASDAQ: AVID) is committed to nurturing the next generation of artists and creators, and supporting equal access in education. Today the company is pleased to announce that its popular Avid Learning Academy program has transitioned to Pro Tools Artist™, providing every student enrolled in the program free access to the new Pro Tools Artist software.
“With Pro Tools Artist in hand, educators can have total confidence that their students will be gaining the skills and experiences that prepare them to fully explore their creativity and career interests,” said Francois Quereuil, Vice President, Product Management for Audio and Music Solutions, Avid. “While Avid Learning Academy now provides a fun and exciting palette with thousands of inspiring sounds and loops, hundreds of plug-ins and even new virtual instruments, aspiring musicians will be excited to know they can create in any genre with the same tools that today’s most popular artists and producers use.”
The Avid Learning Academy program offers high schools a complete curriculum to teach music and media production skills, aligned to the U.S National Core Arts Standards and common CTE pathways in Arts, Media, and Entertainment (AME). Students enrolled in these programs can get a head start on college or their career by learning on the music, TV and film industries’ most popular creative tools and earning pathway certifications with Avid’s industry-recognized professional credentials.  
For music and audio engineering, the Avid Learning Academy program previously included Pro Tools First, a powerful but feature-limited version of Pro Tools. With the release of Pro Tools Artist, Avid is deepening its commitment to empowering the next generation of creators with access to an even more powerful set of music creation tools, continuing to lower the barrier to entry.
Pro Tools Artist is a brand-new subscription offering specifically designed for the millions of next-gen music creators looking to produce great songs, and includes everything needed to make beats, write, record, and mix studio-quality music. And as the Pro Tools session format is the industry’s universal language, they can take their project to any producer or studio around the world to collaborate and take their music to the next level. With hundreds of inspiring instrument sounds and loops and over 100 included plugins, as well as easy-to-use MIDI tools and industry-standard editing and mixing workflows, aspiring music creators now have a professional yet affordable solution to create any style of music.
Avid is working closely with its publishing partner Rowman & Littlefield to update all curriculum and program materials for the 2022-2023 academic year and with ThinkEDU to roll the program out nationwide.  
“This is an exciting update for students who not only benefit from government funding to implement the program, but also now with Pro Tools Artist can create and take their resulting projects to any other Pro Tools system for further editing and finishing,” said Michael Fischler, CEO of ThinkEDU. “The fact that it is free is providing significant value to the Avid Learning Academy program, which positions students for a career in the creative arts on a platform that is ubiquitous in the industry when it comes to live music, studio recording, and post production.”
To enroll or find out more about the Avid Learning Academy program, visit www.avid.com/LearningAcademy.
Avid Powers Greater Creators People who create media for a living become greater creators with Avid’s award-winning technology solutions to make, manage and monetize today’s most celebrated video and audio content—from iconic movies and binge-worthy TV series, to network news and sports, to recorded music and the live stage. What began more than 30 years ago with our invention of nonlinear digital video editing has led to individual artists, creative teams and organizations everywhere subscribing to our powerful tools and collaborating securely in the cloud. We continue to re-imagine the many ways editors, musicians, producers, journalists, and other content creators will bring their stories to life. Discover the possibilities at avid.com in the conversation on social media with the multitude of brilliant creative people who choose Avid for a lifetime of success. 
© 2022 Avid Technology, Inc., Avid, its logo, Pro Tools, Pro Tools Artist, Pro Tools | GrooveCell and Pro Tools | SynthCell are property of Avid. All rights reserved. Other trademarks are property of their respective owners. Product features, specifications, system requirements and availability are subject to change without notice. 
PR ContactsAvid Heather Vosburg[email protected]615.430.3454
Red Lorry Yellow Lorry (Avid’s PR agency)  [email protected]
Primary Logo
Receive full access to all new and archived articles, unlimited portfolio tracking, e-mail alerts, custom newswires and RSS feeds – and more!

source

Panasonic Launches HC-X2 & HC-X20 Camcorders, Plus New Lens – Adorama

Panasonic Launches HC-X2 & HC-X20 Camcorders, Plus New Lens  Adorama
source

Rap beat freestyle – "Massive" l Hip hop instrumentals l Trap beat instrumental 2022

frame:work:london 2022, November 3 -4 At CodeNode – Live Design

source

Avid Hires New Leader for Audio & Music Solutions to Accelerate Subscription Business and … – Bluefield Daily Telegraph

Avid Hires New Leader for Audio & Music Solutions to Accelerate Subscription Business and …  Bluefield Daily Telegraph
source

Tekirdağ Metropolitan Municipality Complex / Lift Studio + One Office – ArchDaily

Tekirdağ Metropolitan Municipality Complex / Lift Studio + One Office  ArchDaily
source

Sweetwater Studios to Host Bainz Hip-Hop Mixing Masterclass – Mix

Update your browser to view this website correctly. Update my browser now
×
Home › Recording›Music Production
Sweetwater Studios will host an in-person Hip Hop Mixing Masterclass with Bainz.
By Mix Staff

Fort Wayne, IN (September 13, 2022)—On Sunday, September 18, 2022, Sweetwater Studios will host an in-person Hip Hop Mixing Masterclass with Bainz, the mix engineer behind top tracks like ‘Ski’ by Young Thug and Gunna, ‘Go Grazy’ by Chris Brown and ‘Franchise’ by Travis Scott, to name only a few.
The Masterclass will bring engineers and producers into Sweetwater Studios’ facilities to learn modern hip-hop and trap music production alongside one of the hottest mix engineers in the industry. Bainz’s work also includes collaborations with Future, Nispey Hussle, Trippie Redd, Mac Miller, Sia, Juice WRLD, Jay Sean, Prateek Kuhad, Rich the Kid, Eric Nam, Shy Glizzy, Offset, Quavo, Machine Gun Kelly, and James Fauntleroy.
The masterclass will take place on Sunday, September 18 from 10am to 5pm, in Sweetwater’s Studio A, utilizing its recently upgraded Rupert Neve Designs 5088 console and hybrid analog/digital workflow. There, Bainz will showcase his workflow and favorite methods to get the most out of hip-hop and trap mixes.
“Over the last decade, Bainz has blown us away with some of our favorite tracks and we’re thrilled to be able to bring him into the studio to share his knowledge,” said Sweetwater Studios producer/engineer Shawn Dealey. “It’s also a great opportunity to demonstrate our versatility at Sweetwater Studios and how you can make the most of a well-equipped studio to deliver an incredible mix, every time.”
The event costs of $495 per attendee; reservations are extremely limited.

© 2022 MIX is part of Future plc, an international media group and leading digital publisher. Visit our corporate site.

source

Today at Elon / Veda Skog ’24 brings harmony to Limelight Records – Today at Elon

By Coral Clark, student
September 13, 2022
Skog was elected president of Limelight Records after two years of dedication to the organization.
Veda Skog ‘24 grew up going to concerts almost every weekend. With a sense of freedom and unbridled joy at every concert, her infectious energy spread to everyone around her.
Music was always an important part of Skog’s journey but she struggled with finding the professional path she wanted to take. Skog, an Odyssey Scholar, came to Elon as an international and global studies major. She desired to make a difference in the world through foreign relations but quickly realized that her passions led back to music.
“The hurdle of having a passion for music was answering the questions of how do I do music and what job would I even be going for,” Skog said. It seems that Skog has found the answers to those questions as she is now president of Limelight Records, Elon University’s sole record label for up-and-coming student musicians.
She got involved with Limelight Records during her first year. But Skog did not want to pursue a music production and recording arts degree at Elon, thinking that Limelight Records would be the outlet for her musical passions.
A COM 1000 course with Tom Nelson, an associate professor of journalism, caused her to venture into the field of communications. Early in her sophomore year, she declared a strategic communications and an outdoor learning and education double major while increasing her involvement in Limelight Records.
Skog’s journey at Limelight Records started when she worked as a manager for a signed artist, Rachel Coughlin, taking headshots to grow Coughlin’s Spotify platform and assisting in any way possible.
In her sophomore year, she was vice president of marketing. She asked to shadow Andrew High, the previous president of Limelight Records, in preparation for her later candidacy. Her shadowing led to a promotion to executive vice president.
Working in a leadership role, Skog was able to oversee the operations of the organization, channel her creativity and observe areas for future progress.
Eliza Spear ‘22 serves as a significant inspiration for Skog in her musical abilities, perseverance and dedication to a career in the music industry. Spear is a recipient of an American Society of Composers, Authors and Publishers Award and is graduated with a major in music and a minor in communications.
She witnessed Skog’s success as a signed Limelight artist and is confident in the growth she will bring to Limelight Records.
“Veda is one of those people that you meet and just know is going to be successful,” Spear said. “She’s absolutely magnetic in everything she does, and Limelight is in the best of hands moving forward. Veda’s leadership, creativity and comfort will take Limelight to new heights.”
She aims to make Limelight Records more community focused and will pull from experiences and lessons learned in Elon Outdoors, the Odyssey Program and the Leadership Fellows program.
“I want to continue to foster, grow and sustain this community as I have done for the past two years,” Skog said. “My experience at Limelight constantly excites me and motivates me to do the work needed to help Limelight be successful.”
Through more live events and community partnerships, Skog plans to counteract the lack of music awareness on campus. She urges students to find needs that aren’t being fulfilled and follow their passions to see what good can lead to.
“If there’s not a space for you, make a space for yourself,” Skog said. “Don’t be afraid to meet other people. That’s where you’re going to make your friends. Be comfortable in the uncomfortable.”
Student Experience

The Kernodle Center for Civic Life announces service opportunities available for students during the fall 2022 semester.
The associate professor of journalism and North Carolina government reporter will co-teach an online, synchronous continuing education class – set to begin Thursday, Oct. 6. – on navigating the media landscape ahead of the November midterm elections.
Elon’s Multifaith Internship has a focus on how experiential learning fosters multifaith understanding.
Skog was elected president of Limelight Records after two years of dedication to the organization.
Get more Elon news delivered to your inbox.


Loading
© 2022 Elon University | All Rights Reserved

source

A fitness expert shares a game plan for living without back pain – CNN

A fitness expert shares a game plan for living without back pain  CNN
source

Mid-Tier Power Index Profile: EQ Accountants – Accountancy Age

David Morrison, partner at EQ Accountants, explains why the firm considers staff development to be its highest priority in seeking to achieve growth
In September, Accountancy Age proudly launched the inaugural edition of the Mid-Tier Power Index (MTPI): a brand new ranking designed to highlight excellence in the accounting industry’s mid-market.
The MTPI seeks to go beyond typical, financial metrics. While Profitability & Growth was taken into account, the entrants were also assessed on Strategic Planning, Professional Excellence, and People & Community.
Securing 22nd place in the ranking was Scotland’s EQ Accountants. The 17-partner firm fared well across all four categories, demonstrating increased focus on talent and service delivery whilst also seeing bottom line growth.
During a conversation with Accountancy Age, David Morrison, partner at EQ Accountants, walks us through the firm’s credentials in each of the judging categories, explaining how a personal, community-based approach to service delivery has allowed the firm to establish a unique identity in a crowded market.

This is probably going to be relevant to all of the sections, but I’d say we are unique among mid-tier firms, running primarily as a general practice. We don’t use separate departments for things like tax, audit, or corporate finance, though we employ many specialists.
What we do is curate a team of people who are highly skilled general practitioners, and we invest heavily in training them across the board. I would put our staff up against any of their peer group in Scotland and say that their tax or audit skills for example would be better than most others.
EQ is all about people doing business with people. In the East of Scotland, which is where we are based, there’s not much of a corporate culture. It’s all about relationships between individuals, and that’s one of the reasons why we focus on general practice as opposed to specialist departments.
To support that, we’ve made sure that every client has three direct points of contact – a Partner, Manager and Controller. If you speak to any potential client, their complaint is typically that they can’t get hold of their accountant. For us, if we haven’t replied to somebody within 24 hours, we’d be really disappointed in ourselves.
The other thing is, when you operate in separate departments, the risk is you don’t get to know the client well enough. But we’ve got people who know their client’s business, their family, and the sector in great detail, and that makes a massive difference for us.
Our strategy is to use that structure to specifically target what we consider to be quality businesses. Each of our offices has a client recruitment committee which identifies and assesses targets, and makes an approach accordingly. What that means is we don’t waste our time on businesses that will never use EQ. So I think we’re very focused.
Training is key for EQ. For example, we have a three-year tax training programme, and the aim of that is to effectively take someone from having no knowledge to being highly adept in tax advisory work. And what’s important is that these people don’t necessarily need to understand the detail of the rules – they need to understand the problem. If you can identify the problem, there’s usually a solution. So our staff are trained with that in mind, and we do a lot of that kind of training on a technical level.
Soft skills are also a really important part of our approach to training, we offer a lot of mentoring around things like professionalism and case studies on how we work with clients. Our aim is to try and allow our staff to grow and develop, and understand why we do things the way that we do them.
We do this because we think it’s really important for people to understand business. You can train anyone to become an accountant, but I don’t think it’s as easy to train someone to be an advisor. So what we’re trying to do is develop people to have a good commercial baseline understanding of how other business people operate.
All of that has a real impact. You can’t legislate for someone leaving the area or wanting a career change, but we rarely lose people to rival local firms, and I think that’s because we’re viewed as an employer of choice in our area, based on the reputation we hold for developing great people.
What we try to do is identify the embodiment of great customer service and apply it to each line of business. I think what that comes down to is identifying the little things that make a big difference in terms of the overall experience of clients.
So for example, instead of asking when the client is going to deliver their accounting information we need, we’ll ask if we can come and collect it. And if the client says they work until 5pm and can’t make themselves available until 6pm, then we’ll meet at 6pm. Our culture is about going above and beyond.
The last few years have obviously been quite challenging operationally, but what we’ve been able to do very quickly is get on the phone with clients and communicate all of the support and guidance that they need. And I think the client goodwill we received from that extra effort has hugely strengthened our client relationships.
Ultimately, you build a business by retaining your clients. So for us, there’s no point in adding new clients if you’re losing them at the other end. I think that’s the mistake many other firms make.
So when we talk about culture, it’s the relationship led service we’re trying to create. When you go into the big cities there’s more of a commodity in the service, but where we are, it’s about people doing business with people, and I can’t stress that enough.
We’ve set ourselves a target of £9.5m for 2022-23, and we’re already ahead of that target, pro rata, despite being not even half way through the year.
I expect our percentage growth may not be as spectacular as other firms, and that’s fine as for us, it’s not just about top line growth., It’s actually the quality of the clients we have that will allow us to grow long-term. If we have a client for 20 years, it’s far better than winning a new client and losing them the next year.
We do set some fairly ambitious targets and we believe we’re going to get close to £10m by March 2023, but it’s just a number. We’re not trying to be biggest firm in the world, or even the biggest in Scotland. We’re trying to be the best.
EQ never sell products. What we deliver is a business relationship.
For us staff recruitment is the number one challenge. We’ve got great people and we think we’ve got the right service model, but adding new people is a constant challenge.
We’ve ramped up our student recruitment significantly in the past couple of years, and we’re doing it again just now for the 2023 intake, but for us it’s critical to identify the right type of person. As I said earlier, we can train anyone to do the technical stuff, but what we need is people with the right mentality.
We will of course face some economic challenges, but because we’re in the East of Scotland, we tend not to hit the highest peaks or the lowest lows economically. But the challenges will certainly hit our clients, and it’s our duty to support them through that changing environment.
And it comes back to the recruitment piece – we need to make sure that we have the right people and enough of them to deliver that tailored service to our clients.
I think what we’re seeing in the profession at the moment is that we’re gradually seeing people automating everything they do and becoming paperless. We’ve invested pretty heavily in being a Platinum Xero partner for example, and we’ve got a lot of people who are very adept with that kind of work. So I think certainly over the next two or three years, it’s an important goal for us to automate everything that our clients are doing.
Beyond that, I think it just comes down to continuing to grow in our key sectors such as agriculture, healthcare and leisure, and trying to recruit the best people, to win the right clients that create opportunities for our people.
Of course we’d love to be a bigger practice by turnover, have more people, operate in more sectors and have more good-quality clients, but we have no ambition to be the ‘number X’ firm in Scotland – it’s just not relevant to us. We’ll grow by being great at what we do, offering tailored, relationship led support and advice.
More about:
In the past decade, the professional services industry has transformed significantly. Digital disruptions, increased competition, and changing market …
In recent months, professional services firms are facing more pressure than ever to deliver value to clients. Often, clients look at the firms own inf…
In a world of instant results and automated workloads, the potential for AP to drive insights and transform results is enormous. But, if you’re still …
The first phase of Making Tax Digital (MTD) saw the requirement for the digital submission of the VAT Return using compliant software. That’s now behi…
Accountancy Age’s new ranking will champion excellence among the industry’s mid-tier Read More…
Just four days remain for UK accounting firms to secure their entries for the brand new industry ranking Read More…
Time is running out for firms to secure their entries for the Accountancy Age’s new ranking Read More…
Entries for the inaugural ranking are open now Read More…
Firms urged to invest in improving company culture and innovation after accounting industry sees record rise in revenues Read More…
Market participants say that workers are increasingly resistant to low pay, insufficient benefits and limited career development Read More…
Accountancy Age spoke with KPMG senior partner for Scotland James Kergon, who outlined his background and the key goals for his tenure Read More…
Accountancy Age spoke with Azets’ new CEO Chris Horne, who outlined the key goals for his tenure Read More…
Copyright © 2022 Accountancy Age

source

Nugen Audio Ships New Halo Vision Plug-In – TV Technology

Nugen Audio Ships New Halo Vision Plug-In  TV Technology
source

TOPSIS aided ensemble of CNN models for screening COVID-19 in chest X-ray images | Scientific Reports – Nature.com

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.
Advertisement
Scientific Reports volume 12, Article number: 15409 (2022)
Metrics details
The novel coronavirus (COVID-19), has undoubtedly imprinted our lives with its deadly impact. Early testing with isolation of the individual is the best possible way to curb the spread of this deadly virus. Computer aided diagnosis (CAD) provides an alternative and cheap option for screening of the said virus. In this paper, we propose a convolution neural network (CNN)-based CAD method for COVID-19 and pneumonia detection from chest X-ray images. We consider three input types for three identical base classifiers. To capture maximum possible complementary features, we consider the original RGB image, Red channel image and the original image stacked with Robert’s edge information. After that we develop an ensemble strategy based on the technique for order preference by similarity to an ideal solution (TOPSIS) to aggregate the outcomes of base classifiers. The overall framework, called TOPCONet, is very light in comparison with standard CNN models in terms of the number of trainable parameters required. TOPCONet achieves state-of-the-art results when evaluated on the three publicly available datasets: (1) IEEE COVID-19 dataset + Kaggle Pneumonia Dataset, (2) Kaggle Radiography dataset and (3) COVIDx.
COVID-19 has a colossal impact on almost every sector of society. Numerous deaths and countless positive cases are causing increased agony. Besides, governments have imposed a lockdown to confine the spread of COVID-19, which largely impacts the world’s economy and culture. More than 590 million positive cases and 6.4 million deaths due to COVID-19 alone have been recorded to date1. Diagnostic techniques, medicines and vaccines are some of the methods greatly researched to save humans from its disastrous consequences2. On the other hand, pneumonia (usually found in children) causes severe respiratory issues that occasionally lead to death. Just like any other disease, in case of both pneumonia and COVID-19, early diagnosis for detection of the same is the most essential step to not only receive proper medical attention but also curb the spread of this disease. One of the most commonly used methods for diagnosis includes a real-time reverse transcription-polymerase chain reaction (RRT-PCR) test from a nasopharyngeal swab sample. However, the major drawback of RRT-PCR is low sensitivity in detecting COVID-19 cases3. Moreover, these pathological diagnosis systems suffer from inter- and intra-observer variability in addition to being time-consuming.
As an alternative to pathological diagnosis, several radiological methods have been explored to diagnose COVID-19 efficiently. Computerised tomography (CT) scans, chest X-rays and magnetic resonance imaging (MRI) are some of the methods used to detect the presence of this deadly virus4. Computer-aided diagnosis (CAD) systems have been beneficial in easing the burden on medical professionals and ruling out the possibility of human errors. Artificial intelligence (AI) based solutions also have proven beneficial5,6. Notably, several research communications can be found recently that detect the presence of these viruses (COVID-19 and pneumonia) from both CT scans and chest X-ray images7. However, the new variants of the virus that causes COVID-19 make detection more challenging. One of the disadvantages of the reliance on radiological methods is the risk caused by radiation, which may damage the molecular structures of the human body8. Despite this risk factor, radiological diagnosis methods are still relied upon as they help in more accurate diagnosis than their counterparts7,9 do.
Convolutional neural network (CNN) aided methods have been successfully used in the past for COVID-19 and pneumonia detection7,10 from the chest X-ray images. The main reasons for preferring them over the duo-feature engineering and classical machine learning approaches are (1) limited need of domain knowledge for extracting learned features from input patterns, (2) capability in handling datasets with imbalanced classes11, and (3) the competence in handling large datasets. CNN-based approaches have been successfully explored in the domains such as speech recognition, medical image classification, handwriting recognition, human action recognition and many more. However, the most common disadvantage of such approaches is that they require high-end infrastructural support to train models properly. Moreover, CNNs are often observed to require assistance from data pre-processing to perform efficiently12.
Ensemble learning, in general, is the process of learning an association of multiple decision making systems. In the context of CNNs, ensemble learning methods aim to capture the complementary information of its constituent base models’ outcomes13. In doing so, many times researchers make use of methods having some learnable parameters. Such techniques require a substantial amount of data to train the classifier. On the other hand, non-trainable ensemble techniques do not need any training data but may require a few parameters to be tuned. For such methods to be used effectively, researchers apply some transformation operations to the raw outputs obtained from the base classifiers which is followed by an aggregation. The aggregation is either a form of hard-voting or soft-voting. In a hard-voting process, each classifier provides a binary class label (votes in favour or against). Whereas in the soft-voting process, each classifier provides a confidence score for each of the classes under consideration, based on which the final classification takes place. Multi criteria decision making (MCDM) algorithms are one of the statistical measure-based techniques used to rank the alternatives based on some given criteria. MCDMs have been extended to a large variety of works such as supplier selection strategy, benchmark machine learning tool selection for COVID-19 diagnosis, ranking of internet of things (IoT) based systems14 and in many more fields. Popular MCDM methods include the technique for order of preference by similarity to ideal solution (TOPSIS)15, vlsekriterijumska optimizacija I ompromisno resenje (VIKOR)16 etc.
Given the above-mentioned facts, in this work, we propose, TOPCONet, a CAD system comprising a lightweight CNN model and a TOPSIS-aided ensemble approach for diagnosis of COVID-19 and pneumonia from chest X-ray images. The lightweight deep CNN model designed here consists of only five convolutional blocks in which each block consists of four layers, namely convolutional, activation, batch normalisation and pooling layers. Each of these layers is used only once for extracting features from images. Subsequently three fully connected layers are added on top of the feature extractor i.e., after the five convolutional blocks. The CNN model is applied on three variants of input chest X-ray images, and thus we obtain three base classification models. The three image variants are: (1) the red channel image (i.e., 1-channel image), (2) normal RGB image (i.e., 3-channel image), and (3) 4-channel image generated by combining edge image obtained by applying the Robert’s edge detection method. It is worth mentioning that the designed CNN-aided classifiers are trained from scratch. Finally, the decisions from the classifiers are aggregated using TOPSIS based ensemble method to obtain the final classification of an inputted chest X-ray image. The proposed TOPCONet model is evaluated on three publicly available standard datasets namely: (1) Kaggle Pneumonia + IEEE COVID dataset, and (2) Kaggle Radiography Dataset and COVIDx, and it achieves performances comparable to state-of-the-art approaches. Concisely, the highlights of the present work are summarised below.
We develop a lightweight but efficient CNN model having much fewer trainable parameters as compared to state-of-the-art CNN models used in CAD systems for the purpose of COVID-19 and pneumonia detection.
We design a TOPSIS, an MCDM technique, aided ensemble method which is found to be more effective than standard ensemble methods such as majority voting, product rule and sum rule.
The performance of the proposed model on three public datasets is comparable to state-of-the-art methods that require heavy computational resources.
The remaining parts of the paper are organised in the following manner. In the “Literature review” section, we first briefly review some past works from the literature of COVID-19 and pneumonia diagnosis system and then analyse each of them to describe their advantages and shortcomings. In the “Preliminaries” section, we briefly describe some prerequisites useful in describing the working procedure of the proposed model. Next, in the “Proposed method” section, we describe in detail how our model (i.e., TOPCONet) works. In the “Experimental results” section, we present all the experimental findings related to this work while we analyse the results and explain each of the experimental findings in the “Discussion” section. We also summarise the advantages and limitations of our work in this section. In the “Conclusion” section, we end with concluding remarks and some scopes for future research.
Several research attempts were made by researchers to design effective CAD systems for COVID-19. Here, we review some of the existing works employed for early detection of COVID-19 and/or pneumonia from chest X-ray images. Wang et al.17 designed a CNN architecture, dubbed as COVID-Net, that achieved 80% sensitivity for COVID-19 detection. This is one of the fundamental works that laid the ground-work for further research in this domain with deep learning-based CAD models for detection of the said diseases. The article by Demir12 presented a long short term memory (LSTM) network-based approach where the Sobel edge detection and marker-controlled watershed segmentation approach were used in the pre-processing stage. The author was able to significantly boost the model performance, but the computational cost was considerably high because of training both CNN and LSTM models. Convolutional LSTM (ConvLSTM) based approach has also been seen in the literature recently18,19.
Classical handcrafted feature engineering-based approaches have also proven their competence for COVID-19 detection. Panetta et al.20 proposed a novel shape-dependent Fibonacci-p patterns-based feature descriptor for COVID-19 detection. In another work, Chandra et al.21 first extracted the gray-level cooccurrence matrix (GLCM) and histogram oriented gradient (HOG) based features and then used the binary grey wolf optimizer (BGWO) based feature selection technique to select the near-optimal feature set. Finally, a bi-stage majority voting-aided ensemble mechanism was utilised with the seven different machine learning-based classifiers. The authors classifies normal and abnormal images quite successfully, but the model performance was not good while classifying infected samples into pneumonia and COVID-19 cases. One of the possible reasons for such a behaviour might be related to a recent medical research findings22 where the authors state “In some cases, definite discrimination of the two (COVID-19 and pneumonia) entities might be impossible solely based on the imaging, however, some radiologic features may suggest one diagnosis over the other”.
Oh et al.23 used a patch-based classification strategy to mitigate the difficulties due to a limited training sample while using a CNN model-based classification strategy. They first segmented the images using deep CNNs and then generated random patches to train the CNNs using the ResNet-18 model. Thus they were able to train the CNN model using limited training datasets. The approach also lowered the training cost by a margin. In the said work, the normalised validation accuracies were used as fuzzy measures for the base classifiers but it may not work with other datasets. Similarly, in a very recent article, Paul et al.24 proposed to use an inverted bell-shaped weighted average scheme to ensemble the probabilistic outcomes of deep transfer learning models. In another work, Das et al.25 designed a bi-stage CAD system for the detection of COVID-19 and pneumonia-infected chest X-ray images. In the first stage, the authors segregated the infected chest X-ray images from the normal ones, and in the later stage, the COVID-19 cases were identified from the infected cases. The authors used the pre-trained VGG-19 as the feature extractor, and a shallow learner was employed as the classifier. The performance of the bi-stage model was good but once a sample is classified erroneously in the first stage, it remained unclassified throughout. Moreover, the bi-stage classification scheme for detecting a COVID-19 or pneumonia case needs classification of a sample twice, which lengthens the testing time. A stacked CNN model using sub-models of VGG-19 and XceptionNet was proposed by Gour et al.26, but use of such large CNN models like VGG-19 and XceptionNet increases the training and inference time. Hasoon et al.27 used a combination of image preprocessing techniques and machine learning algorithms to get 6 different models those are local binary pattern (LBP)-kNN, HOG-kNN, Haralick-kNN, LBP-SVM, HOG-SVM, and Haralick-SVM for chest X-ray image classification. The authors showed that the use of machine learning algorithms failed to produce better result over the CNN based models. A new COVID-19 detection model CoWarriorNet was proposed by Roy et al.28 that consists of two networks a classification network and a confidence network. The proposed architecture is a derivative of U-Net architecture, which results in poor extraction of image-derived information.
Ouchicha et al.29 provided a CNN architecture with a sort of inter-network skip connections for screening of COVID-19 cases from chest X-ray images. The authors used two parallel CNN models and connected each of the layers of networks with inter-network and intra-network residual connections for feature transfer to encounter the vanishing gradient problem. Additionally, it is interesting to note that downsampling and skip connections have also been proposed in the literature in the past for COVID-19 detection30. Khuzani et al.31 employed multiple feature extraction schemes in which a novel pooling strategy was used to extract only the relevant features. Finally, the authors used a fully connected neural network to classify chest X-ray images using the extracted features. The authors used the machine learning-based approach, instead of the conventional CNN approach which could significantly reduce the computational cost, but the approach failed to tackle the class imbalance issue. In another work, Kenaway et al.32 first used a popular CNN architecture as a feature extractor and then proposed an advanced version of the squirrel search optimisation algorithm to select the best set of features. They finally used the multi-layer perceptron (MLP) on top of the selected features for classification, leading to a significant increase in training time, as the transfer learning model is fine-tuned. Besides, the feature selection approach is trainable and the classifier also needs time to train. Gour et al.33 introduced a uncertainty aware CNN model known as the UA-ConvNet model. The problem of this CNN model is that in its initial stage, a pre-trained EfficientNet-B3 model is fine-tuned, which makes it time-taking. Bashar et al.34 used transfer learning concept while using CNN models like AlexNet, GoogleNet, VGG-16, VGG-19 and DenseNet to classify chest X-ray images. The authors enhanced the input image quality using techniques like anisotropic diffusion filter, Fourier transform, and edge-aware local contrast manipulation. Here, again the use of transfer learning techniques makes the process computationally expensive to train. Similar problem faced by Senan et al.35, where ResNet50 and AlexNet models were used to generate features which are combined with GLCM and LBP features.
Khan et al.36 proposed an XceptionNet-based classification model, which performed well, but the metrics reveal that it somewhat failed to handle imbalanced classes despite using a CNN-based architecture. Hussain et al.37 proposed a 22-layers deep CNN model. Goel et al.38 proposed a CNN model whose hyperparameters were optimised using a nature-inspired meta-heuristic approach: BGWO. This method provided a good alternative to the tiring grid search method, but the model has a slightly higher false-positive rate, which might not be useful in practical field. Aslan et al.39 designed a hybrid approach using a modified version of the AlexNet hybridised with bi-directional LSTM (BiLSTM) providing better results worth-while results in processing the features for classification. Multi-level feature extraction technique was used by Naeem et al.40 where the authors first extracted features like GIST, scale invariant feature transform (SIFT), and CNN based deep features and then they used LSTM network to perform classification. Similarly, classical classifiers like artificial neural network (ANN), SVM, KNN, and deep learning classifiers like recurrent neural network (RNN), LSTM were used for predicting COVID-19 cases from chest X-ray images by Goyal et al.41. The problem with LSTM and RNN architectures are that they are prone to over-fitting and difficult to apply dropout algorithm there.
Numerous deep learning-based methods have been explored in the past for designing CAD systems for the diagnosis of COVID-19 and pneumonia from chest X-ray images. Most of them aimed at designing a system that could perform more efficiently than its predecessors, which is the primary objective of designing a new CAD system at its initial stage. Hence, in many cases, these models became computationally heavier. However, an ideal CAD system should be computationally inexpensive and memory-efficient for its wide applicability and portability so that such a model can be easily executable in a resource-constrained environment without compromising its performance. Besides, the increment of learnable parameters in CNN models increases the space and time requirements. Another important issue is that such models at times extract some redundant features which in turn consume additional space and time during execution without much improvement of performance42. Additionally, according to the lottery ticket hypothesis43, we can prune the learnable parameters of a heavy CNN model to make it much more cost-efficient and portable without compromising its performance. Keeping these facts in mind, we aim at design a lightweight CNN model that can detect COVID-19 and pneumonia-infected chest X-ray images from normal ones. Additionally, it is to be noted that the authors of the works6,21,44 used a hierarchical approach to design better CAD models for diagnosis of the said diseases, in which they applied various heavier CNN architectures at each stage. The use of multiple pre-trained CNN models along with evaluating an input image several times makes these models inefficient in terms of storage and computation needs. Hence, we evaluate an input image only once to make it more memory and time-efficient. Apart from these, some methods6,45 used ensemble techniques in which different CNN models were applied as base classifiers to obtain different learned features from input images. However, the study46 shows that we can gather complementary information by pre-processing the original image to extract features. Hence in our model, we pass two differently pre-processed images along with the original one to the CNN model for extracting three different features from it. We also observe from the literature that the work reported by Paul et al.45 used a machine learning-based ensemble technique and showed its effectiveness in improving the end performance. In a work by Pramanik et al.6, where the authors propose a fuzzy aggregator to ensemble the CNN outcomes which is non-trainable in nature.
In this work, to capture the complementary information provided by the base classifiers, we propose an ensemble with an MCDM-based aggregator known as TOPSIS. The ensemble framework has a few tunable parameters compared to the machine learning-based ensemble techniques. The use of such methods aims at developing a lightweight but effective CAD model for distinguishing COVID-19 and pneumonia-infected chest X-ray images from normal ones.
The Robert’s edge detection technique uses 2D spatial gradient calculation to detect edges in an input image. It uses two operators of size (2times 2), known as Robert cross operators, which help in calculating the gradient image. The gradient image is used for identifying the edges in the image. Here, our idea is to extract the minute details present in a chest X-ray image and thus use a small-sized mask while calculating gradient image, which would be helpful rather than using larger masks. This is why we chose this edge detection technique. The Robert cross operators are shown in Eqs. (1) and (2).
In this edge detection technique, an input image is convolved with these two operators i.e., (delta _x) and (delta _y). Let (delta _x) and (delta _y) be convolved with a gray-scale image (say, (I_g)) and generate two gradient images (say, (G_x) and (G_y)) using Eqs. (3) and (4).
In Eqs. (3) and (4), (‘*’) is the convolutional operator. The magnitude image corresponding to the gradient images (say, G) is calculated using Eq. (5).
We obtain the final edge image using a threshold value (say, th). In the present work, we use the mean of the gradient magnitude values appearing in G. The parameter th is calculated using Eq. (6).
In Eq. (6), M and N stand for the height and width of the input image respectively. Finally the edge image (say, (I_e)) is obtained using Eq. (7).
This is probably one of the most revolutionary research findings in the domain of document image processing47, which later extended to multiple fields such as video processing48, speech processing and emotion recognition. CNN is used as a discriminative architecture49 inspired by the concept of a time delay neural network (TDNN). In TDNN, the weights are shared in a temporal dimension, which ultimately reduces computation. On the other hand, the CNN model does not require weights to be shared, and convolutions are replaced by matrix multiplication as in standard NNs. This replacement significantly decreases number of weights. Moreover images can be fed as raw input to the network, whereas in classical NN models, features, extracted from the images, are fed to the NNs. In other words, CNNs are seen as a combination of a learned feature extractor and a classifier. Weights for both the classifier and the feature extractor model are learned using the back propagation algorithm47. Some important parts/characteristics of any CNN models are described in the later parts.
One of the most important parts of any CNN model is the convolutional layer which is suggested by its name. In a convolutional layer, a linear operation occurs in which the input image is convolved with a set of learnable weights. For applying the convolutional process on an image, a two-dimensional array (i.e., matrix) of weights is used which is known as the kernel or filter. The filter is smaller in size compared to the input image. To perform convolution element-wise matrix multiplication is performed between the filter matrix and a filter-sized patch of the input image. After this, all the values of the patch are added together to produce a single value. By using a filter smaller than the input size it can then be applied to the various parts of the input during convolution. Thus, the same weights can be applied to the various parts of the input data resulting in translation invariance. The invariance to translation is an important property for knowing whether a feature is present in the input image rather than in the location of the feature in the image.
Activation functions play a crucial role during the training of the CNN models. Let (z_i=g(x,w_i)) be the mathematical formulation of any neuron and (y_i=f(z_i)) be the value that is to be transmitted to the next neuron. The use of an activation function is popular because it helps in countering the vanishing gradient problem associated with any CNN model. In standard NNs, activation functions control the flow of information among the neurons present therein. There are several diverse activation functions in literature at present. The choice of activation functions mainly depends on the objective of the work at hand. Probably, one of the most popularly used activation functions is the rectified linear unit (ReLU) which is well used for image recognition-based tasks47.
Initially suggested by Ioffe et al.50, batch normalisation greatly improvises the convergence of the CNN model during training and also helps produce stable results. Batch normalisation helps in coordinating the update of weights in multiple layers in the model. To do so, it scales the output of the layer such as standardising the activations of each input variable or the activations of the nodes of the previous layer. Here, standardising means rescaling data to have a mean of 0 and a standard deviation of 1. This standardisation stabilises during the training phase of the CNN model and also speeds up the training time.
A pooling layer is designed to reduce the dimension of feature maps. This layer ensures that the relevant information passes on to the next level of the network while the redundant features are discarded. Thereby pooling serves as a filter mechanism for the extracted features. In this work, we use max-pooling which returns the maximum local value of each pooling window.
In fully connected network (FCN) layers, the neurons of the current layer remain connected to each neuron of the previous layer. Figure 1 shows how a single neuron of one layer connects with multiple neurons of another layer during the construction of a multi-layered FCN model. In this figure, the node y is represented with the weighted sum of its inputs as shown in Eq. (8). In this equation, the parameters w and b are optimised using the back propagation algorithm. Relevant studies in the past have shown that a neuron having a larger w value means that it is more important. The parameter b is the bias variable used to shift the value of y. The output layer of the FCN part of a CNN model is an N-dimensional vector, where N is the number of classes for a classification problem. A pictorial description of the classical multi-layered FCN model is given in Fig. 2.
A pictorial representation of how a neuron of one layer connects with neurons of other layer.
A pictorial description of classical feed-forward neural network with one input layer of size 12, two hidden layers each of size 8 and an output vector of length 4. This figure has been generated using a tool developed by Alex LeNail https://alexlenail.me/NN-SVG/.
The final FCN layer uses a softmax function (see Eq. 9) that determines the class level probabilities of an input for a classification problem.
In Eq. (9), (z=[z_1,z_2,…,z_N]) is the input to the softmax function.
TOPSIS, proposed by Hwang and Yoon15, is a simple (though effective) multi-criteria decision analysis (MCDA) or MCDM method. It is derived from the idea that the geometric distance of the highest-ranked alternative is the nearest to the ideal solution while the lowest-ranked alternative is the nearest to the worst solution. The positive ideal solution capitalises on the beneficial attributes most while minimising on the cost attributes. The TOPSIS method uses all of the information in the form of a decision matrix and criteria weights as a vector and generates a ranking for all the alternative candidate. A detailed procedure is described in Algorithm 1.
In this work, we propose TOPCONet, an ensemble-based approach using lightweight CNN based three base learners (hereafter termed Classifier 1, Classifier 2 and Classifier 3) to detect COVID-19 and pneumonia from chest X-ray images. The base classifiers use the same CNN architecture but work with different forms of input images. The image variants are red channel image (i.e., 1-channel red image), original RGB image, and a 4-channel image prepared by combining the edge image generated by the Robert’s edge detection method with the original RGB image used in Classifier 1, Classifier 2 and Classifier 3 respectively. To begin with we resize all the images to the spatial dimension of (224times 224) pixels to feed to the base classifiers. While evaluating we estimate the class probability scores using the base classifiers and feed them to the TOPSIS-based ensemble method. The complete pipeline for our proposed model can be found in Fig. 3.
The pipeline for the proposed model. Here, the CNN architecture for all three base classifiers is the same. Images are taken from the public datasets found in51,52.
It has been described earlier that a CNN architecture consists of two main components viz., a feature extractor and a classifier network. Therefore, while designing a customised CNN architecture for any specific task, both the feature extractor and the classifier should be stressed upon. A very deep network consisting of many layers might be effective for extracting very deep features but on the other hand, optimisation of the model becomes a difficult task which means the optimisation of the very deep network is much difficult when compared to small-sized networks. The same has been empirically and theoretically shown by Choromanska et al.53. The major motivations for using CNNs as a feature extractor over handcrafted features is the presence of shared weights, pooling, local connections and many possible layers47. Therefore for each of these layers, while designing, one thing that should be considered is the extraction of the best possible features for the classifier to perform efficiently. It must also be kept in mind to minimise the number of redundant features. To efficiently deal with this, nowadays deep feature selection-based approaches are also gaining popularity42,54. Filter pruning strategy is one of the appproaches to reducing the number of redundant features; one such work was communicated by Luo et al.55. On a similar note, in this work, we aim to extract meaningful and less redundant features. Considering the mentioned facts, the customised CNN designed with five blocks of convolutions, is carefully tuned to extract the best possible features and minimize the redundant extracted features. Each level of convolution is followed by a layer of activation to control the flow of information to the next level of convolution and for, batch normalisation for re-scaling and re-centring of the values to have smoother learning and max-pooling to get rid of redundant features. It the end, an FCN with two hidden layers is coupled with exponential linear unit (ELU) activation function to impart a non-linearity to the input signal. Thus, we ensure that both the feature extractor and the classifier work efficiently. The detailed structure is shown in Fig. 4.
Architecture of the proposed CNN model.
As already been mentioned, in the present work, we employ a TOPSIS-aided ensemble method to decide the final class of an input chest X-ray image depending on the class level confidence scores generated by the base classifiers. We use the available classes as the alternatives and the confidence scores from each of the base classifiers as the criteria values for each of the alternatives (i.e., classes). The ijth element of the decision matrix (i.e., (r_{ij})) used in the TOPSIS method is constructed by considering the confidence score of the ith base classifier corresponding to the jth class. We employ all the steps as described in Algorithm 1 on the decision matrix to obtain the top most-ranked alternative which here is the final class. It should be noted that we resort to column normalisation of the decision matrix to comply with the TOPSIS algorithm. During the normalisation process, we divide the input score by the square root of the summation of squares of the observations. In our case, all the criteria are beneficial.
As the proposed ensemble method is non-trainable, it is important to discuss the computational complexity of the model. In this part, we discuss the computational complexity using the ({mathscr {O}})-notation. The ensemble method requires constant number of operations for each of the samples. For each of n base classifiers and m alternatives (i.e., number of classes), normalisation and weight assignment. The complexity of the model is ({mathscr {O}} (ntimes m)). The complexity while determining the positive and negative ideal solution and distance from the two is ({mathscr {O}} (ntimes m)). Furthermore, the complexity of finding the top ranked alternative is ({mathscr {O}}(m)).
All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki Declaration and its later amendments.
In this section, we discuss and also attempt to explain the experimental findings. It should be noted that all the performances are evaluated on machines supported by Nvidia Tesla T4 GPUs. The programming environment used to train and test our model is Python 3.8.
To evaluate the performance of our proposed method for COVID-19 and pneumonia detection from chest X-ray images we use samples from two datasets. The first dataset (termed as Dataset-1 hereafter) consists of chest X-ray images of pneumonia-affected patients and normal subjects from Guangzhou Medical Centre: Chest-X-ray-Pneumonia52 and COVID-19 affected patients from IEEE COVID-chest X-ray-dataset51. The second dataset, termed here after as Dataset-2, is hosted publicly on Kaggle and is popularly known as the COVID-19 Radiography Database56. Dataset-2 has been created in collaboration of researchers from Qatar University (Qatar), University of Dhaka (Bangladesh), Pakistan and Malaysia. The dataset consists of chest X-ray images of COVID-19 positive, normal and pneumonia-infected cases. The distributions (i.e., number of samples in train and test set) of both the datasets have been shown in Table 1. It is to be mentioned here that during model training we have randomly selected 15% of the training samples as validation samples.
In typical CNN models, hyperparameter tuning is a very essential and tedious job. Hyperparameters of a deep learning model to an extent indicate the learning capability of such a model. It is essential to set two main hyperparameters, namely batch size and learning rate to obtain near-optimal model performance. The batch size controls the number of samples to be loaded for training at one go while the learning rate determines the learning capability of the model which says to what extent the newly acquired information overrides the previous. Here we estimate the hyperparameters using samples of Dataset-1 and propagate these parameters to train the model for Dataset-2. Here we would also like to mention that we tune the said hyperparameters only for Classifier 2 as it takes the standard RGB images as input. For tuning the hyperparameters, we randomly select 15% of the training dataset as the validation samples. For this, we employ the grid search method to select the near-optimal batch size and the learning rate. The learning rate is selected from the set ({1e-02,1e-03,1e-04,1e-05,1e-06}) while for selecting batch size we use the set ({8,16,32}). From Fig. 5 it is observed that the validation accuracies converge to their optimum solution for learning rate = (1{text {e}}-4) and batch size = 16. For smooth learning we rely on the popularly used Adam optimiser and categorical cross-entropy as the loss function.
Validation accuracy (in %) with respect to different learning rates used to train the customised CNN model.
It has already been mentioned that we propose a customised CNN model that is lightweight in nature (see Fig. 4). Based on this CNN model, three different classifiers are formed. We compare the performances of the proposed classifiers (including TOPCONet) with state-of-the-art CNN models following the transfer learning concept and training from the scratch. In the transfer learning approach, we load the pre-trained weights of the respective CNN models which are obtained after training the models on the ImageNet dataset. Subsequently, using the loaded weights we fine-tune the models on both datasets. To maintain uniformity during performance comparison, the total number of epochs used is 15 for both state-of-the-art CNN models and the proposed classifiers. All the results are evaluated on test samples of the said datasets i.e., following hold-out test set samples approach. The comparative results are presented in Tables 2 and 3 for Dataset-1 and Dataset-2 respectively.
From these tables the number of trainable parameters for the proposed Classifier 1, Classifier 2 and Classifier 3 can be observed as 306,467; 339,235; and 355,619, respectively. Classifier 3 uses the maximum number of trainable parameters (355,619) which is 28.49% of trainable parameters used in the SqueezeNet model (1,248,424), which itself is considered to have the least number of trainable parameters among the well-known deep CNN models such as MobileNetV2 (3,504,872), VGG-19 (143,667,240) and DenseNet121 (7,978,856). Moreover, these tables also show that the total number of trainable parameters used in the proposed TOPCONet model is 1,001,324 (trainable parameters of Classifier 1 + Classifier 2 + Classifier 3 + TOPSIS-aided ensemble) which is evidence that although TOPCONet consists of three base classifiers it uses less number of trainable parameters than the state-of-the-art CNN models. Hence, we can safely comment that the proposed TOPCONet is computationally less expensive than the popularly used state-of-the-art CNN models used in literature in terms of trainable parameters.
For Dataset-1, when compared in terms of accuracy, the proposed Classifier 3 among the three base classifiers demonstrates the highest recognition accuracy which is 0.58% more than the second-highest 98.20% obtained using the VGG-19 model which uses transfer learning concept. In the case of the precision score, the MobileNetV2 model trained from scratch on Dataset-1 offers a better result (98.39%) which is 0.05%, more than our Classifier 2. For recall the DenseNet121 (trained from scratch) and MobileNetV2 (following transfer learning concept) give better results over our base classifiers. In the case of the F1-score, the proposed Classifier 3 performs the best. However, TOPCONet outperforms all the CNN models on this dataset in terms of all the parameters. Now comparing our base classifiers with other popularly used CNN models on Dataset-2, it is evident that the proposed Classifier 3 outperforms all the other popular CNN models in terms of recognition accuracy with a difference of (0.63%) with respect to the second-best result given by the present Classifier 2. It is also to be noted that apart from the accuracy, the Classifier 3 performs better in terms of precision, recall and F1-score values which are (0.54%), (0.48%) and (0.77%) better than the second-best method.
Although in some cases the proposed base classifiers are not able to outperform the other CNN models, it is to be kept in mind that it has far fewer of trainable parameters than those other CNN models which makes it easier to train, making it appropriate use in a resource constraint environment. It is evident from the training time column that the proposed base classifiers take far lesser time to train than the other state-of-the-art CNN-based classifiers in the case of both the datasets. In the case of CNN models, training from scratch takes more time than using the transfer learning approach. However, in the case of the proposed classifiers although they are being trained from scratch they take less time than the standard CNN models using transfer learning approach.
In a nutshell, from the experimental results discussed in this subsection and shown in Tables 2 and 3 we can safely comment that (1) the proposed customised CNN model takes much less time to train from scratch as compared to state-of-the-art CNN models (please note Classifier 2 take the same input type (i.e., RGB image) as other state-of-the-art CNN models), (2) our customised CNN model also takes less time as compared to other CNN models when they are trained according to the transfer learning concept, (3) both the customised CNN model-based base classifiers and TOPCONet have less trainable parameters compared to other CNN models, (4) unsurprisingly, TOPCONet outperforms all the models with moderate training and testing time.
We implemented some standard ensemble methods such as product rule, sum rule, weighted average method and majority voting method. After evaluation on the test sets of Dataset-1 and Dataset-2, the performances of the ensemble methods are shown in Table 4. Here we would like to mention that for implementing the TOPSIS-aided ensemble approach, we have first ranked the base classifiers depending on validation accuracy and then assigned the weights (termed as criteria weights in the TOPSIS method) 0.55, 0.35 and 0.1 chronologically from the best to the worst-performing base classifiers. We set these values experimentally. This parameter tuning is conducted only for Dataset-1 while using the same values for evaluating TOPCONet on Dataset-2. In terms of accuracy, the TOPSIS aided ensemble method outperforms all the other standard ensemble methods used here for comparison on Dataset-1. The difference in accuracy is (0.16%) when compared to the second best performers: weighted average and sum rule based ensemble methods. Considering the results on Dataset-2, the difference in accuracy between the highest accuracy obtained using the proposed method and the second best result obtained by product rule ensemble method is (0.19%).
In terms of precision for Dataset-1, the sum rule and weighted average-based ensemble methods exceed that of the proposed method by (0.18%) and for Dataset-2, the product rule-based ensemble outperforms the proposed TOPSIS-aided ensemble method by (0.26%). The highest result for performance in terms of the recall value is given by the proposed method which is (98.26%) for Dataset-1 but for Dataset-2 it is the product rule based ensemble that exceeds the proposed method by (0.12%). If we compare in terms of F1-score on the Dataset-1, it can be concluded that the majority voting ensemble method gives the best F1-score with (0.05%) more than the proposed method while for Dataset-2, it is the product rule based ensemble method that gives the highest F1-score with the difference being (0.19%) from the proposed method.
Moreover, the performances of the ensemble methods of weighted average and sum rule are almost the same. It may have happened because, in the case of the weighted average method, we take the normalised validation accuracies as the weights for each classifier which themselves are almost similar and so despite using weighted average ensemble to prioritise the best performing classifiers, the ensemble gives almost equal priority to the three classifiers which in turn become similar to sum rule based-ensemble method.
In this section, we describe the execution time for different steps in TOPCONet with the machine configuration described previously. The time required to process one batch of samples (i.e., 16 images at a time) at different stages is tabulated in Table 5. From Table 5 it is clear that the total time required to test one sample using the proposed approach is ((240+12+13+14+4)/16) which is nearly 18 ms. The increase in training time for Classifier 2 and Classifier 3 over Classifier 1 is attributed to the fact that the number of trainable parameters for Classifiers 2 and 3 is higher than Classifier 1. This is because Classifier 2 and Classifier 3 accept 3-channel and 4-channel input images respectively, while Classifier 1 accepts 1-channel input images.
To test the effectiveness of the proposed model, we also evaluate our model performance using the standard 5-fold cross-validation scheme. In this setting, we split the dataset into five separate folds. Each time sample of four splits is considered a training samples while the rest samples are used for the testing. The detailed experimental outcomes are recorded in Table 6. From Table 6 it is evident that for Dataset-1, the TOPCONet model performs the best on Fold 4 experiment with accuracy 98.24%. In the case of Dataset-2, the highest accuracy for the proposed ensemble method is achieved during Fold 3 experimentation which is 97.85%. For the base classifiers in the case of Dataset-1, the highest accuracy for Classifier 1 is achieved in Fold 2, for Classifier 2 it is Fold 4 and for Classifier 3 it is Fold 3 experiment. For Dataset-2, during Fold 3 experimentation we attain the highest accuracy for all the base classifiers.
To test the effectiveness of the proposed TOPCONet model, we compare its performance with the performances of some state-of-the-art methods proposed by Khan et al.36, Jain et al.57, Hussain et al.37, Ismael et al.58, Das et al.25, Goel et al.38, and Paul et al.24 on Dataset-1 while for Dataset-2 we compare the methods proposed by Aslan et al.39, Ouchicha et al.29, Kedia et al.59, Ahmad et al.60, Chowdhury et al.56, Sedik et al.18, Wu et al.30, Panetta et al.20, Yang et al.61, Paul et al.24, Gour et al.33, Gour et al.26, Hasoon et al.27, Bashar et al.34, Senan et al.35, Naeem et al.40, Goyal et al.41, Roy et al.28 and Senan et al.35. To have a fair comparison with these methods include the experimental setups, especially the approach followed to prepare the samples of train and test sets, used by these methods. We found that two major approaches were used by these researchers: (1) partitioning the samples of the entire dataset into train and test sets18,20,25,34,35,36,38,39,57,58,59,61, as shown in Table 1 (we call this experimental setup as hold-out test set approach) and (2) standard 5-fold cross validation approach26,27,29,30,33,37,40,41,56,60 as described in subsection “Performance of TOPCONet model using 5-fold cross validation technique”. However, the method designed by Paul et al.24 do not follow any of the mentioned experimental setups and hence we have developed the models at our end and evaluated following our experimental schemes. For the method proposed by Paul et al.24 we use transfer learning concept. Likewise, we also estimate the performances of the proposed TOPCONet model using both the mentioned approaches. On the other hand, the method proposed by Bashar et al.34 and Senan et al.35 designed for 4-class classification problem. Hence, for the comparison with our method we train and test these two models on the present datasets using their experimental setups, and consider the best performance obtained. The comparative results are recorded in Table 7 and Table 8 for Dataset-1 and Dataset-2 respectively.
For Dataset-1, among the state-of-the-art methods evaluated following the hold-out test set approach, the methods proposed by Senan et al.35 and Das et al.25 stand first and second with a test accuracy of 98.70% and 98.45% respectively. Also, the best recall, F1-score and precision values are given by the optimized ResNet50 model by Senan et al.35 with values 98..67%, 98.67% and 98.00% respectively. Whereas, the pre-trained Xception model proposed by Jain et al.57 gives the best precision of 98.00% among the state-of-the-art methods. Now, when comparing TOPCONet with state-of-the-art methods that followed the hold-out test set based experimental setup, results from Table 7 are the evidence that the highest accuracy for the Dataset-1 is given by the proposed method, which is (0.08%) more than the optimized ResNet50 model proposed by Senan et al.35. In terms of precision score, our method exceeds the pre-trained Xception model by Jain et al.57 and the optimized ResNet50 model by Senan et al.35 by (0.67%). However, in terms of F1-score the optimized ResNet50 model outperforms our TOPCONet method by ((0.33%).) and in terms of recall it is outperformed by Senan et al.35 by (0.67%). In the case of methods that followed the five-fold cross-validation approach, the proposed method outperforms the state-of-the-art method that employed CoroDet.
For Dataset-2 among the state-of-the-art methods that followed five-fold cross-validation experimental setup, the best performer is the UA-ConvNet model which is proposed by Gour et al.33 with an accuracy of (98.90%) while the second-best accuracy is provided by the ensemble of InceptionV3 and MobileNetV2 proposed by Ahmad et al.60 with a value of (98.77%). In the case of precision, recall and F1-score the ensemble of InceptionV3 and MobileNetV260 outperforms all other methods in the case of 5-fold cross-validation experimental setup. It also outperforms the proposed method with 5-Fold cross-validation by (0.57%), (0.55%), (0.56%) and (0.92%) in precision, recall, F1-score and recognition accuracy respectively.
When it comes to the comparison of the methods using the hold-out data splitting method, the best accuracy, precision, recall and F1-score are achieved by the hybrid architecture using the modified AlexNet model and BiLSTM with the value of (98.70%), (98.77%), (98.76%) and (98.76%) respectively. However, this method outperforms the proposed method in precision, recall, F1-score and accuracy by (0.93%), (0.91%), (0.91%) and (0.09%) respectively. Although from the comparison, it can be observed that some methods have outperformed the proposed method in Dataset-2 for both the hold-out test set and five-fold cross-validation experimental setups but it is to be kept in mind that the number of parameters used by the proposed method is far less than that of these state-of-the-art methods. The total number of parameters of the combination of three classifiers is around 1 million which is far less than that of AlexNet, MobileNetV2 and InceptionV3 models as shown in Table 2.
To investigate scalability and robustness of the proposed TOPCONet model, we have evaluated its performance on a recently published large daraset, known as COVIDx CXR-3 Dataset62. This dataset was made public by Wang et al.17. It consists of 30,000 chest X-ray images from over 16,490 subjects (either COVID-19 positive patients or normal subjects). The images in the dataset are already partitioned into train and test sets by the researchers in the work17. In the train set of the dataset, there are 16,400 chest X-rays of COVID-19 positive patients and 13,992 chest X-rays belonging to non COVID-19 subjects whereas, the test set contains 200 chest X-ray images of the COVID-19 positive and negative patients each. We have followed the standard dataset divisions to conduct our experiments. The performance of the TOPCONet along with the base classifiers are provided in Table 9. In order to compare the performances of our model on this dataset, we compare it with some state-of-the-art COVID-19 detection methods like Xception model used by Jain et al.57, the CheXNet model by Chowdhury et al.56, the ResNet50 model used Senan et al.35, the CovidConvLSTM model by Dey et al.19 and the use of VGG-16 as a transfer learning model that was used by Bashar et al.34. Performances of these state-of-the-art models are also recorded in Table 9 along with performances of our final model and base classifiers. From Table 9, it can be concluded that the proposed TOPCONet model gives the highest performance in all metrics like accuracy, precision, recall and F1-score. The second and third highest results in all metrics are given by Classifier 1 and Classifier 2 respectively. However, the Classifier 3 is outperformed in terms of accuracy by other state-of-the-art methods. Apart from the performance, present models acquire less trainable parameters than the past methods used here for comparison.
It has already been mentioned that in this work, we propose a new CAD system (i.e., TOPCONet) for COVID-19 and pneumonia detection from chest X-ray images. The TOPCONet model is computationally inexpensive and efficient. The performance of the model is evaluated on two publicly available datasets and the obtained results are described in the previous section. However, in this section, we discuss more findings of the TOPCONet model. To start with we discuss the performance and the number of training parameters of all the basic components of the model. It is to be noted here that TOPCONet comprises of four major components: three base classifiers, namely Classifier 1, Classifier 2 and Classifier 3 and TOPSIS aided ensemble process. The performances on Dataset-1 and Dataset-2 using the experimental setup are discussed in the subsection “Dataset description” and the number of trainable/tunable parameters present in these major components are provided in Table 2. From the performances recorded in Tables 2 and 3 we can see that Classifier 3 that accepts 4-channel input images performs the best among the base classifiers for both datasets. From this observation, we can safely comment that combining the edge image, generated by Robert’s edge detection technique, with the original image helps us to obtain better performance. From these results, we can also say that our model learns the association of features.
In addition to these, we can also observe that after the ensemble of the outputs from the three base classifiers using the TOPSIS-aided ensemble method, it results in an increase of accuracy by (0.25%) for Dataset-1 and in the case of Dataset-2 the increase in accuracy is (0.33%). The results establish that the TOPSIS-aided ensemble method can learn the better association of confidence scores provided by the classifiers. For a deeper understanding, we examine the confusion matrices of the base classifiers and the TOPSIS aided ensemble method for all the datasets. The confusion matrices are shown in Fig. 6. From the figure, we can conlcude the TOPSIS aided ensemble works optimally for all the three datasets. Hence we can also claim that the proposed method works well across the datasets. The proposed ensemble method is able to capture multiple complimentary information provided by the base classifiers for COVID-19 and pneumonia detection.
Confusion matrices obtained on Dataset-1, Dataset-2 and COVIDx dataset.
We can also observe from Tables 2 and 3 that a certain number of parameters are trained for each of the base classifiers. This can be explained because the input dimension of each of the base classifiers differs. Moreover, it must be noted that the ensemble approach consists of several criteria (in this case, three). Corresponding to each of the criteria, a criteria weight is assigned. These weights are manually tuned i.e., these are not trained by any learning methods.
For further analysis of the proposed 3-channel CNN model (i.e., Classifier 2), we use grad-cam analysis, which works by using the gradients generated inside the model to create a heat-map to show the regions where the model focuses to perform the intended classification task. In Fig. 7, the original chest X-ray images along with their grad-cam counterparts are shown. In this figures, we observe that the TOPCONet focuses mainly on the lung area. It is noteworthy that a few activation regions can be seen outside the lung. This is probably because the model learns a few redundant features along with the useful ones. These redundant features are the ones that get insignificant importance from the classifier (FCNNs in this case) during the final classification.
Grad-cam images along with original chest X-ray images for the base classifier 2 trained on the Dataset-1. Images are taken from the public datasets found in51,52.
The advantages of the proposed model are listed below:
The proposed CNN model can learn effective features from input images of different types.
It has fewer tuning parameters compared to state-of-the-art CNN models.
The TOPSIS aided ensemble method can efficiently aggregate the confidence scores estimated by the base classifiers to provide the final class.
The ensemble method used here has very few tunable parameters (same as the number of base classifiers used in ensemble mechanism) and hence very little additional cost is required.
To provide our observations on the proposed work, we also mention some of its limitations which are as follows:
The proposed CNN model might not work well on rotated and arbitrarily cropped images. It might require different augmented training samples like other CNN models.
The TOPSIS aided ensemble method fails to classify a sample if two of the base classifiers incorrectly classify it with high confidence.
As we have used deep learning models as our base learners, it demands resources such as graphics processing unit (GPU), dedicated random access memory (RAM) etc. for training the same.
Keeping in mind the present scenario of COVID-19 cases, the situation needs to be addressed properly to help humanity to come out of this crisis as soon as possible. In the present work, we propose a CAD system, called TOPCONet, for screening COVID-19 and pneumonia which performs well on publicly available three chest X-ray datasets. In TOPCONet, we first pre-process the images to obtain different forms of the original images, which are further used to train and evaluate the model. The skeletal structure of the base classifiers of the ensemble approach is the same, however, the type of input differs for each of the cases. We combine the decisions of these base classifiers using the proposed TOPSIS aided ensemble method. The proposed model can learn deep features better, and it performs at par with the state-of-the-art methods. It is noteworthy to mention that our customised CNN model uses a very less number of trainable parameters compared to the popularly used CNN models in several deep learning-aided CAD systems. It should be noted that these datasets suffer from the problem of imbalanced classes. Hence further works can be focused on handling this issue efficiently. Since our model gives an end-to-end classification result for COVID-19 and pneumonia diseases using chest X-ray images and is lighter in nature, it may be employed in detecting such diseases to ease the burden on radiologists and doctors. However, the proper advice of medical professionals should be considered along with the outcome from this model. Apart from these, further work may be focused on using domain knowledge as it may increase the model performance. Last but not the least, the present model can be used to extract features first and then a suitable feature selection technique may be employed to select a near-optimal subset of features that may further reduce the storage and time requirements.
WHO Coronavirus (COVID-19) Dashboard|WHO Coronavirus (COVID-19) Dashboard With Vaccination Data. https://covid19.who.int/. (Accessed 25 August 2022).
Ella, R. et al. Efficacy, safety, and lot-to-lot immunogenicity of an inactivated SARS-CoV-2 vaccine (BBV152): Interim results of a randomised, double-blind, controlled, phase 3 trial. Lancet. https://doi.org/10.1016/S0140-6736(21)02000-6 (2021).
Article  PubMed  PubMed Central  Google Scholar 
Fang, Y. et al. Sensitivity of chest CT for COVID-19: Comparison to RT-PCR. Radiology. 296, E115–E117 (2020).
Article  Google Scholar 
Kevadiya, B. D. et al. Diagnostics for SARS-CoV-2 infections. Nat. Mater. 20, 593–605. https://doi.org/10.1038/s41563-020-00906-z (2021).
ADS  CAS  Article  PubMed  PubMed Central  Google Scholar 
Santosh, K. AI-driven tools for coronavirus outbreak: Need of active learning and cross-population train/test models on multitudinal/multimodal data. J. Med. Syst. 44, 1–5 (2020).
Article  Google Scholar 
Pramanik, R. et al. A fuzzy distance-based ensemble of deep models for cervical cancer detection. Computer Methods Programs Biomed. 219, 106776. https://doi.org/10.1016/j.cmpb.2022.106776 (2022).
Article  Google Scholar 
Santosh, K. & Ghosh, S. COVID-19 imaging tools: How big data is big?. J. Med. Syst. 45, 1–8 (2021).
Article  Google Scholar 
Risks of Radiation|UCSF Radiology. https://radiology.ucsf.edu/patient-care/patient-safety/radiation-safety/risks-of-radiation. (Accessed 06 August 2021).
Ai, T. et al. Correlation of chest CT and RT-PCR testing for coronavirus disease 2019 (COVID-19) in China: A report of 1014 cases. Radiology. 296, E32–E40. https://doi.org/10.1148/radiol.2020200642 (2020).
CAS  Article  PubMed  Google Scholar 
Mukherjee, H. et al. Deep neural network to detect COVID-19: One architecture for both CT Scans and Chest X-rays. Appl. Intell. 51, 2777–2789 (2021).
Article  Google Scholar 
Buda, M., Maki, A. & Mazurowski, M. A. A systematic study of the class imbalance problem in convolutional neural networks. Neural Netw. 106, 249–259. https://doi.org/10.1016/j.neunet.2018.07.011 (2018).
Article  PubMed  Google Scholar 
Demir, F. DeepCoroNet: A deep LSTM approach for automated detection of COVID-19 cases from chest X-ray images. Appl. Soft Comput. 103, 107160. https://doi.org/10.1016/j.asoc.2021.107160 (2021).
Article  PubMed  PubMed Central  Google Scholar 
Tao, Z., Liu, H., Li, S., Ding, Z. & Fu, Y. Marginalized multiview ensemble clustering. IEEE Trans. Neural Netw. Learn. Syst. 31, 600–611 (2019).
MathSciNet  Article  Google Scholar 
Mohammed, M. A. et al. Benchmarking methodology for selection of optimal COVID-19 diagnostic model based on entropy and TOPSIS methods. IEEE Access 8, 99115–99131. https://doi.org/10.1109/ACCESS.2020.2995597 (2020).
Article  Google Scholar 
Yoon, K. P. & Hwang, C.-L. Multiple Attribute Decision Making: An Introduction (Sage Publications, 1995).
Book  Google Scholar 
Opricovic, S. Multicriteria optimization of civil engineering systems. Faculty Civ. Eng. Belgrade 2, 5–21 (1998).
Google Scholar 
Wang, L., Lin, Z. Q. & Wong, A. COVID-Net: A tailored deep convolutional neural network design for detection of COVID-19 cases from chest X-ray images. Sci. Rep. 10, 19549. https://doi.org/10.1038/s41598-020-76550-z (2020).
ADS  CAS  Article  PubMed  PubMed Central  Google Scholar 
Sedik, A. et al. Efficient deep learning approach for augmented detection of Coronavirus disease. Neural Computing and Applications, 34(14), 11423–11440 (2022).
Article  Google Scholar 
Dey, S., Bhattacharya, R., Malakar, S., Schwenker, F. & Sarkar, R. CovidConvLSTM: A fuzzy ensemble model for COVID-19 detection from chest X-rays. Expert Syst. Appl. 206, 117812 (2022).
Article  Google Scholar 
Panetta, K., Sanghavi, F., Agaian, S. & Madan, N. Automated detection of COVID-19 cases on radiographs using shape-dependent Fibonacci-p patterns. IEEE J. Biomed. Health Inform. 25, 1852–1863. https://doi.org/10.1109/JBHI.2021.3069798 (2021).
Article  PubMed  Google Scholar 
Chandra, T. B., Verma, K., Singh, B. K., Jain, D. & Netam, S. S. Coronavirus disease (COVID-19) detection in Chest X-ray images using majority voting based classifier ensemble. Expert Syst. Appl. 165, 113909. https://doi.org/10.1016/j.eswa.2020.113909 (2021).
Article  PubMed  Google Scholar 
Zarei, F., Reza, J., Sefidbakht, S., Iranpour, P. & Haghighi, R. R. Aspiration pneumonia or COVID-19 infection: A diagnostic challenge. Acad. Radiol. 27, 1046 (2020).
Article  Google Scholar 
Oh, Y., Park, S. & Ye, J. C. Deep learning COVID-19 features on CXR using limited training data sets. IEEE Trans. Med. Imaging 39, 2688–2700. https://doi.org/10.1109/TMI.2020.2993291 (2020).
Article  PubMed  Google Scholar 
Paul, A., Basu, A., Mahmud, M., Kaiser, M. S. & Sarkar, R. Inverted bell-curve-based ensemble of deep learning models for detection of COVID-19 from chest X-rays. Neural Comput. Appl.https://doi.org/10.1007/s00521-021-06737-6 (2022).
Article  PubMed  PubMed Central  Google Scholar 
Das, S., Roy, S. D., Malakar, S., Velásquez, J. D. & Sarkar, R. Bi-level prediction model for screening COVID-19 patients using chest X-ray images. Big Data Res. 25, 100233. https://doi.org/10.1016/j.bdr.2021.100233 (2021).
Article  Google Scholar 
Gour, M. & Jain, S. Automated COVID-19 detection from X-ray and CT images with stacked ensemble convolutional neural network. Biocybern. Biomed. Eng. 42, 27–41 (2022).
Article  Google Scholar 
Hasoon, J. N. et al. COVID-19 anomaly detection and classification method based on supervised machine learning of chest X-ray images. Results Phys. 31, 105045 (2021).
Article  Google Scholar 
Roy, I., Shai, R., Ghosh, A., Bej, A. & Pati, S. K. CoWarriorNet: A novel deep-learning framework for COVID-19 detection from chest X-ray images. New Generation Comput. 1–25 (2021).
Ouchicha, C., Ammor, O. & Meknassi, M. CVDNet: A novel deep learning architecture for detection of coronavirus (COVID-19) from chest X-ray images. Chaos Solitons Fractals 140, 110245. https://doi.org/10.1016/j.chaos.2020.110245 (2020).
MathSciNet  Article  PubMed  PubMed Central  MATH  Google Scholar 
Wu, T., Tang, C., Xu, M., Hong, N. & Lei, Z. ULNet for the detection of coronavirus (COVID-19) from chest X-ray images. Comput. Biol. Med. 137, 104834 (2021).
CAS  Article  Google Scholar 
Zargari Khuzani, A., Heidari, M. & Shariati, S. A. COVID-Classifier: An automated machine learning model to assist in the diagnosis of COVID-19 infection in chest X-ray images. Sci. Rep. 11, 9887. https://doi.org/10.1038/s41598-021-88807-2 (2021).
ADS  CAS  Article  PubMed  PubMed Central  Google Scholar 
El-Kenawy, E.-S.M. et al. Advanced meta-heuristics, convolutional neural networks, and feature selectors for efficient COVID-19 X-ray chest image classification. IEEE Access 9, 36019–36037. https://doi.org/10.1109/ACCESS.2021.3061058 (2021).
Article  PubMed  Google Scholar 
Gour, M. & Jain, S. Uncertainty-aware convolutional neural network for COVID-19 X-ray images classification. Comput. Biol. Med. 140, 105047 (2022).
CAS  Article  Google Scholar 
Bashar, A., Latif, G., Ben Brahim, G., Mohammad, N. & Alghazo, J. COVID-19 pneumonia detection using optimized deep learning techniques. Diagnostics 11, 1972 (2021).
CAS  Article  Google Scholar 
Senan, E. M., Alzahrani, A., Alzahrani, M. Y., Alsharif, N. & Aldhyani, T. H. Automated diagnosis of chest X-ray for early detection of COVID-19 disease. Comput. Math. Methods Med. 2021 (2021).
Khan, A. I., Shah, J. L. & Bhat, M. M. CoroNet: A deep neural network for detection and diagnosis of COVID-19 from chest X-ray images. Comput. Methods Programs Biomed. 196, 105581 (2020).
Article  Google Scholar 
Hussain, E. et al. CoroDet: A deep learning based classification for COVID-19 detection using chest X-ray images. Chaos Solitons Fractals 142, 110495 (2021).
MathSciNet  Article  Google Scholar 
Goel, T., Murugan, R., Mirjalili, S. & Chakrabartty, D. K. OptCoNet: An optimized convolutional neural network for an automatic diagnosis of COVID-19. Appl. Intell. 51, 1351–1366 (2021).
Article  Google Scholar 
Aslan, M. F., Unlersen, M. F., Sabanci, K. & Durdu, A. CNN-based transfer learning-BiLSTM network: A novel approach for COVID-19 infection detection. Appl. Soft Comput. 98, 106912. https://doi.org/10.1016/j.asoc.2020.106912 (2021).
Article  PubMed  Google Scholar 
Naeem, H. & Bin-Salem, A. A. A CNN-LSTM network with multi-level feature extraction-based approach for automated detection of coronavirus from CT scan and X-ray images. Appl. Soft Comput. 113, 107918 (2021).
Article  Google Scholar 
Goyal, S. & Singh, R. Detection and classification of lung diseases for pneumonia and COVID-19 using machine and deep learning techniques. J. Ambient Intell. Human. Comput.. 1–21 (2021).
Das, S., Chatterjee, A., Dey, S., Saha, S. & Malakar, S. Breast cancer detection from histology images using deep feature selection. in Proceedings of International Conference on Frontiers in Computing and Systems, 323–330 (Springer, 2023).
Frankle, J. & Carbin, M. The lottery ticket hypothesis: Finding sparse, trainable neural networks. in International Conference on Learning Representations (2018).
Pereira, R. M., Bertolini, D., Teixeira, L. O., Silla, C. N. Jr. & Costa, Y. M. COVID-19 identification in chest X-ray images on flat and hierarchical classification scenarios. Comput. Methods Programs Biomed. 194, 105532 (2020).
Article  Google Scholar 
Paul, A., Pramanik, R., Malakar, S. & Sarkar, R. An ensemble of deep transfer learning models for handwritten music symbol recognition. Neural Comput. Appl. 34, 10409–10427. https://doi.org/10.1007/s00521-021-06629-9 (2022).
Article  Google Scholar 
Mukhopadhyay, S., Hossain, S., Malakar, S., Cuevas, E. & Sarkar, R. Image contrast improvement through a metaheuristic scheme. Soft Comput. 1–20 (2022).
LeCun, Y., Bengio, Y. & Hinton, G. Deep learning. Nature 521, 436–444. https://doi.org/10.1038/nature14539 (2015).
ADS  CAS  Article  PubMed  Google Scholar 
Banerjee, A., Singh, P. K. & Sarkar, R. Fuzzy integral-based CNN classifier fusion for 3D skeleton action recognition. IEEE Trans. Circ. Syst. Video Technol. 31, 2206–2216. https://doi.org/10.1109/TCSVT.2020.3019293 (2021).
Article  Google Scholar 
Arel, I., Rose, D. C. & Karnowski, T. P. Research frontier: Deep machine learning-a new frontier in artificial intelligence research. IEEE Comput. Intell. Mag. 5, 13–18. https://doi.org/10.1109/MCI.2010.938364 (2010).
Article  Google Scholar 
Ioffe, S. & Szegedy, C. Batch normalization: Accelerating deep network training by reducing internal covariate shift. in Proceedings of the 32nd International Conference on Machine Learning, (Bach, F. & Blei, D. eds.)vol. 37 of Proceedings of Machine Learning Research, 448–456 (PMLR, Lille, France, 2015).
Cohen, J. P., Morrison, P. & Dao, L. COVID-19 image data collection. arXiv 2003.11597 https://github.com/ieee8023/covid-chestxray-dataset (2020).
Monney, P. Chest X-ray Images (Pneumonia)|Kaggle. https://www.kaggle.com/paultimothymooney/chest-xray-pneumonia/version/2. (Accessed 08 February 2021).
Choromanska, A., Henaff, M., Mathieu, M., Arous, G. B. & LeCun, Y. The loss surfaces of multilayer networks. in Artificial Intelligence and Statistics, 192–204 (PMLR, 2015).
Bandyopadhyay, R., Basu, A., Cuevas, E. & Sarkar, R. Harris Hawks optimisation with Simulated Annealing as a deep feature selection method for screening of COVID-19 CT-scans. Appl. Soft Comput. 111, 107698. https://doi.org/10.1016/j.asoc.2021.107698 (2021).
Article  PubMed  PubMed Central  Google Scholar 
Luo, J.-H., Wu, J. & Lin, W. Thinet: A filter level pruning method for deep neural network compression. in Proceedings of the IEEE International Conference on Computer Vision, 5058–5066 (2017).
Chowdhury, M. E. et al. Can AI help in screening viral and COVID-19 pneumonia?. IEEE Access 8, 132665–132676 (2020).
Article  Google Scholar 
Jain, R., Gupta, M., Taneja, S. & Hemanth, D. J. Deep learning based detection and analysis of COVID-19 on chest X-ray images. Appl. Intell. 51, 1690–1700 (2021).
Article  Google Scholar 
Ismael, A. M. & Şengür, A. Deep learning approaches for COVID-19 detection based on chest X-ray images. Expert Syst. Appl. 164, 114054 (2021).
Article  Google Scholar 
Kedia, P. et al. CoVNet-19: A Deep Learning model for the detection and analysis of COVID-19 patients. Appl. Soft Comput. 104, 107184 (2021).
Article  Google Scholar 
Ahmad, F., Khan, M. U. G. & Javed, K. Deep learning model for distinguishing novel coronavirus from other chest related infections in X-ray images. Comput. Biol. Med. 134, 104401 (2021).
CAS  Article  Google Scholar 
Yang, D. et al. Detection and analysis of COVID-19 in medical images using deep learning techniques. Sci. Rep. 11, 1–13 (2021).
ADS  Article  Google Scholar 
Zhao, A. COVIDx CXR-2: Chest X-ray images for the detection of COVID-19. https://www.kaggle.com/andyczhao/covidx-cxr2. (Accessed 28 December 2021).
Download references
We are sincerely obliged to the Center for Microprocessor Applications for Training Education and Research (CMATER) research laboratory of the Department of Computer Science and Engineering, Jadavpur University, Kolkata, India for providing infrastructural support to this work.
Department of Computer Science and Engineering, Jadavpur University, Kolkata, 700032, India
Rishav Pramanik & Ram Sarkar
Department of Electrical Engineering, Jadavpur University, Kolkata, 700032, India
Subhrajit Dey
Department of Computer Science, Asutosh College, Kolkata, 700026, India
Samir Malakar
Centre for Artificial Intelligence Research and Optimisation, Torrens University Australia, Fortitude Valley, Brisbane, QLD, 4006, Australia
Seyedali Mirjalili
Yonsei Frontier Lab, Yonsei University, Seoul, Republic of Korea
Seyedali Mirjalili
You can also search for this author in PubMed Google Scholar
You can also search for this author in PubMed Google Scholar
You can also search for this author in PubMed Google Scholar
You can also search for this author in PubMed Google Scholar
You can also search for this author in PubMed Google Scholar
R.P., S.D., Sa.M. and R.S. conceived the idea and analysed the supporting theory. R.P. designed the experimental protocol. R.P. and S.D. conducted the experiments. Sa.M., Se.M. and R.S. analysed the implications of the outcomes. Se.M. and R.S. supervised the entire work. All authors reviewed and approved the final manuscript.
Correspondence to Seyedali Mirjalili.
The authors declare no competing interests.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
Reprints and Permissions
Pramanik, R., Dey, S., Malakar, S. et al. TOPSIS aided ensemble of CNN models for screening COVID-19 in chest X-ray images. Sci Rep 12, 15409 (2022). https://doi.org/10.1038/s41598-022-18463-7
Download citation
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s41598-022-18463-7
Anyone you share the following link with will be able to read this content:
Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.
Advertisement
Advanced search
Scientific Reports (Sci Rep) ISSN 2045-2322 (online)
© 2022 Springer Nature Limited
Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

source

Uber introduces audio-recording safety feature – The Mail & Guardian – Mail and Guardian

Subscribe
Subscribe
Uber is piloting an audio-recording security feature in Johannesburg and Pretoria after successful launches in Latin American and the US in 2019.
Chinese company Didi was the first in South Africa to offer in-app audio recordings on its ride-hailing service but it exited the market after just a year in April 2022. It introduced the feature in China in 2018.
Uber Sub-Saharan Africa general manager Kagiso Khaole told the Mail & Guardian the company had wanted to make sure the audio-recording feature adhered to the law. 
“When you build it with privacy in mind, there’s a lot of consideration that needs to be put in place.
“We wanted to not only meet the legal requirements, but needed to meet Uber’s standards, and this is why it has taken us this long to bring this feature to South Africa,” said Khaole.
The additional security feature is opt-in, enabling either the rider or driver to record audio at the beginning of the trip, or at any point in its duration, by pressing the shield icon on the app and selecting “record audio”.
Both parties can see the availability of the feature but neither know when it is being used. The audio clips are saved in one-minute increments that are time-stamped, in case a phone loses power. It will stay on a user’s device for 15 days, after which it will be automatically deleted.
Importantly, you’re not able to access it, said Khaole. It’s encrypted so neither the rider, the driver nor Uber is able to listen to it. 
“It’s only once you file a safety incident, that’s when we will be able to unencrypt it. There’s a safety support team that has the credentials to be able to unencrypt the file once it is shared.”
The audio recording feature went live in Johannesburg and Pretoria on Thursday, 8 September, and will be rolled out to the rest of South Africa within a few weeks. The company is investigating how to launch it in the other sub-Saharan markets in which it operates. 
Khaole said it would give peace of mind to users of the platform and it couldn’t be done without adhering to privacy principles.
The audio-recording pilot in South Africa is an addition to the other safety features available in the sub-Saharan region, including RideCheck, trusted contacts and PIN verification.
At the media event in Johannesburg, head of new verticals at Uber Eats for sub-Saharan Africa Cikida Gcali-Mabusela also touched on the company’s plan to become a “virtual mall of the future”.
“We’ve got the restaurants and now we’re coming for everything else,” said Gcali-Mabusela. Other categories Uber has ventured into include
convenience stores, pharmacy outlets, alcohol sales and retail, “which shows us that the possibilities are endless”. 
The company has added grocery chains, fuel companies and textbook delivery services in the region. In South Africa, it has partnered with Exclusive Books and Game. Customers also have the option to buy anything from perfume to traditional herbs from Uber Eats.
Gcali-Mabusela said they were working on a product-replacement feature, for when items were out of stock, to give the customer the option to either get a refund or to choose a “best match” from other products. The feature is expected to roll out at the end of the year.
Another shopping feature to look out for is scheduling. 
“It will come in the hour that you want and you can track it. With other e-commerce experiences, you’re never really sure when it’s coming,” said Gcali-Mabusela. 
All liquor purchases on the UberEats app will have an identity-verification feature that scans the barcode of a South African ID before completing the purchase.
The company said the future of logistics and technology in emerging markets looked promising.  
If this story helped you navigate your world, subscribe to the M&G today for just R30 for the first three months. 
Subscribers get access to all our best journalism, subscriber-only newsletters, events and a weekly cryptic crossword.”

Nafisa Akabor is a freelance technology journalist.






M&G Premium




For all the latest information and advice from the Department of Health relating to Covid-19, please visit sacoronavirus.co.za

 

Blow The Whistle
Contact Us
Story Tip-offs
Privacy Policy
Cookie Policy
Ethics & Social Media Policy
About Us
Corrections And Clarifications
Advertising
Terms of Service and Security
Mail & Guardian Jobs
Find property for sale
We value your feedback
Subscriber login FAQ
Subscriptions
All material © Mail & Guardian Online. Material may not be published or reproduced in any form without prior written permission.

source

E48 – Sound Mixing a Short Film Pt. 1 – Adobe Premiere Pro CC 2020

Annalibera on opening for Pavement, their latest album 'Moonbath' – Iowa Public Radio

I interviewed Annalibera’s Anna Gebhardt in Greenwood Park in Des Moines during a hailstorm. Taking sanctuary in the open-air shelter by the pond, we read the graffiti scrawled all over its wooden posts.
“Don’t get good at sumthin u don’t like doing.”
Our conversation that day ranged from Terry Pratchett’s satirical fantasy series Discworld and Olaf Stapledon’s speculative history of the cosmos Star Maker to NBA teams, to silent film music (the subject of her college thesis).
Gebhardt is an adventurous and restless artist who emerged from a classical music background in 2015 with two debuts: the shoegazey Nevermind I Love You, where Annalibera is a simmering alt-rock trio, and the noise cassette loveil, where Annalibera is a screeching bionic animal made up of Gebhardt and her laptop.
Her follow-up, the self-produced Opia (2018), where Annalibera is a flamboyant studio operation, burst out in just about every direction—dream pop, chamber pop, jazz ballads, new wave, drone rock, impressionistic jams—all anchored by Gebhardt’s expressive voice.
Annalibera’s latest album Moon Bath was made in collaboration with producer Philip Rabalais, previously a member of Iowa bands Trouble Lights and Utopia Park.
Moon Bath again plunges Gebhardt into colorful new territory: the world of electronic pop, with detours into sound collage and electronica. Moon Bath was released on CD by vibrant independent label Nova Labs right before the 2020 lockdown. Swallowed by the pandemic and unsupported by a tour, the album didn’t receive digital release until now.
This September, Annalibera will play for their biggest crowds yet as the opener for Pavement’s reunion shows in Denver, Kansas City, and Saint Paul. Iowans who want to catch the new iteration of the project can see them at the Maximum Ames Music Festival earlier that week, on Saturday, Sept. 17. The band’s current lineup is Anna Gebhardt, Ryan Stier, Caleb Swank-Ferrara, and Trent Derby, and will feature Gebhardt on sax.
Our conversation follows below, condensed and edited for clarity.
NK: In a way Moon Bath continues the adventures in electronic music that Annalibera began with loveil. Did you know Moon Bath was going to be an electronic pop album when you were writing it? Is that why you started working with producer Philip Rabalais?
AG: I didn’t set out to write it that way specifically. Some ideas just needed that treatment or sort of come like that. […] With Phil, Trouble Lights was playing at Wooly’s, and I chased him down when he was heading to his car after the show. I told him that I had some songs I wanted to record and that, with his style, he would be the perfect person to work with. He said yes. And it was a really nice experience making Moon Bath with him. It’s my favorite [experience making a record] so far.
I actually saw Phil at the first concert I went to when I moved here in 2008. I drove over to Iowa City, to Gabe’s. I went to see Tilly and the Wall and this band opened for them called Porno Galactica which was Phil and his brother Dom. They were doing computer music and hanging from the rafters.
Hanging from the rafters?
Literally. They jumped up there and climbed around on them, hanging down and singing and stuff. It was awesome, I mean, it was really good.
Were you into electronic music at the time?
Yeah, yeah. I grew up listening to late 90s, early 2000s Madonna, for example. I was just listening to the Scissor Sisters driving here. You know the Scissor Sisters?
I don’t think so.
Well, I mean…[plays “Filthy/Gorgeous” from her phone] And I loved Eiffel 65 as a kid. I listened to that album so much, I could sing the whole thing exactly.
Moon Bath is more pop-inspired than your previous albums, with danceable beats and a few big choruses, but it also resists some pop tendencies with musical left turns and traces of Annalibera’s experimental side. What’s your relationship to pop music in general?
I like pop music, whenever people ask me what kind of music I play, I mention pop. But it’s hard to get the point across since “pop” means a lot of different styles. […] I really like making stuff that feels very balanced, like it already existed. Almost basic. That’s pop music I think.
“Kikyo II” is one of my favorite songs on the album. Is it a sequel to loveil’s “Kikyo”? It interpolates a vocal chant from the earlier song.
Yes, “Kikyo II” is about the same character that I wrote the first one about. We’re doing that one on the tour.
Is there a connection between loveil and your classical music background? You jumped right into the world of noise music with that album but everything in it is well structured and it has the kind of composition you might hear in a 20th-century classical work.
Oh cool. I was actually reading a book about [16th century Italian composer] Gesualdo at the time. He was a noble, a creepy Italian dude in his creepy castle. Killed his wife. He made this insane weird music, very emotional, very intense. I sent loveil to the guy who wrote [the Gesualdo book I was reading] and he was like, “I see no connection between Gesualdo and your music.” But he was nice, he said thanks for sharing.
Do you prefer making music by yourself or with other people? Moon Bath and Nevermind I Love You both had a collaborative process while Opia and loveil seem more self-driven.
I have a choral background, I grew up doing music in church…I don’t know what to call that feeling of community or transcendence when you’re making music with people. I feel good playing piano by myself too;I love doing that. But once I got out in the world, off the farm—as a kid I couldn’t really leave the farm unless my parents let me, so piano was really big for me at that time in my life—[I was more excited about making music with other people than] playing by myself. But I like both.
When I started making music on my own, I was trying to get out of the jail of classical music. Opera didn’t want me. I noticed people don’t seem inspired when I sing that, but they do seem inspired when I sing indie or Americana styles.
Was opera something you wanted to do?
Not really, it was just what the program focused on. I had a great time– I mean, I had an intense time but also a great time with the choral program at Drake. We were making music at a really high level. I had some amazing moments performing, transcendent I-know-everything-about-the-universe moments. Have you had those? It’s like, “why can’t I remember that?” when they’re over. “I knew all the answers! Where’d they go?” I’ve always tried to find that making my own music. Never found it so far really. I still sing with choirs, I’ve pretty consistently sung with choirs since college.
It seems like you’re interested in other arts too. You directed a film for loveil which references filmmaker Jean Luc Godard’s Two or Three Things I Know About Her early on. Your song “Sarah Lucas” is dedicated to the artist…
I’m going to make more films; I really like it. I like Kenneth Anger, Maya Deren, that’s the kind of thing I’d want to make more of. I’m always writing and planning films and scenes.
The “Renaissance man” isn’t really a thing these days. You have to be such a specialist. I think it’s easy to get in a tunnel vision with whatever your thing is and then be like, “Nothing else matters.” I love visual art, I love film, I try to see live theater as much as I can, I love going to operas, I love sports. I really like watching the NBA. I think sports are an art, I get a lot out of them. […] During the pandemic I decided I have a new big dream. I want to be a comic writer– as in funny, not graphic novels.
That makes sense. There’s a humor to some your songs, especially on Opia. Or “Love In A Recession” which, when I first heard it at one of your shows, I thought was a cover of an old John Prine country song or something. It has that kind of sensibility. It starts out kind of funny and then gets more emotional as it goes on and as the performance becomes more intense.
Thanks. That’s why with Nevermind I Love You…I have a lot of distance from it now and I’m proud of it when I listen to it, but at the time it didn’t turn own how I envisioned it. It’s not funny at all, all of the lightness isn’t there. It’s very serious. But there are a lot of people who like that about it and that’s important to me too, the conversation that happens after you share what you’ve made.
You mentioned going on tour in Europe around the time of Nevermind I Love You.
We had started playing with this woman Anne. She was going to go study in Stockholm and said we should go on a European tour. So we did. We played loveil material pretty much. I booked the tour myself and the routing was really, really bad. We’d be like, “Okay, today we’re driving across Germany and then we’re gonna drive back across Germany again for the next set.” You can’t do that. Whatever, we did it and it was insane. It’s a really long story, I could probably write a book about it. And that’s the only way I can really tell the story, unless we’re in, like, a fifty-hour car ride together. I don’t even like telling stories. There are stories and they could stand alone but it’s like, you don’t understand the effect of what I’m saying because for two weeks straight before this, the same amount of intensity of insane stuff was happening back to back the whole time. So me just telling this standalone part of it doesn’t really get the point across.
But anyway, I got to do this cool last hurrah with her. Her being in the band helped me a lot. Actually the European tour did too. People think the music industry is maybe [half women], and it does kind of seem like that because there are a lot of frontwomen. But the lead singer is just a small part of the music industry. There’s all the players, there’s the bartenders, the promoters—everyone involved is a dude, basically. That tour was the first time I realized that when we do soundchecks and there’s dudes in my band, no one talks to me. So when me and Anne would do soundchecks, the engineers would say, “Who am I supposed to talk to?” Like, one of us! There’s only two of us, and we’re both women so you’ve gotta take your pick, man, sorry!
How do you play loveil songs live?
When I play that material live, I’ll play solo usually or maybe with a guitarist. One time I played a show with Bob [Nastanovich of Pavement, at Baby’s All Right in Brooklyn], and he was on drums. I use several pedals and a microphone. It doesn’t really sound like the album; I mess with the sounds and play improvisational versions of the songs. I’m kind of chopping up and mutating the material with a sampler.
What’s your personal history with Pavement?
When I was in high school my friend would put their songs on mix CDs for me. “We Dance” was a really big song for me then. I was listening to “Cut Your Hair” yesterday and was like, “perfect.” I never thought I’d get to see them live, let alone open for them.

source

In Memoriam Ron Logan – The UCLA Herb Alpert School of Music – UCLA Herb Alpert School of Music

In the summer of 1961, Ron Logan traveled with the UCLA band for a tour of Europe. The senior trumpet major led a breakout jazz band that played for King Frederick of Denmark. In attendance as a special guest of the king was Walt Disney.
It turned out to be an omen. Logan, who graduated from UCLA with bachelor’s and master’s degrees in Music and Music Education, began working for the Walt Disney Company in 1960, playing trumpet. He worked elsewhere as well, both as a musician and an educator. Logan served as director of bands at James Monroe High School in Sepulveda before becoming the director of bands and jazz studies at Long Beach City College, which he held until 1978.
Then he went to work for Walt Disney full time. He began as music director at Walt Disney World in Florida, conducting the park’s marching band. He rose to Executive Vice President and Executive Producer of Walt Disney Entertainment Worldwide.
Logan produced large-scale live entertainment for Disney’s parks, including Fantasmic, which played at Disneyland in 1992 and Disney’s Hollywood Studios in 1998, Indiana Jones Epic Stunt Spectacular!, which played at Disney’s Hollywood Studios in 1989, Festival of the Lion King, which played at Disney’s Animal Kingdom Theme Park in 1998, and Buffalo Bill’s Wildwest Show which played at Disneyland Paris in 1992. He was part of the production team that brought Beauty and the Beast to Broadway, and he helped found Walt Disney Theatrical Productions, where he served as president.
He also produced five Superbowl halftime shows and, by the end of his career, oversaw all live entertainment at Disney’s theme parks, on its cruise ships, and in its resorts.
Logan retired in 2001. He was named a Disney legend in 2007. He passed away on August 30, 2022 at the age of 84.
The School of Music honors Ron Logan’s career and work, from his accomplishments as a trumpet performance major at UCLA, to directing a high school marching band, to producing so much live entertainment. He brought joy to countless people.
© 2022 UCLA Herb Alpert School of Music
Site Design by Kley, Inc.

source

Live at Leeds: how Jesse Marsch brought New York grit to the Premier League fight – The Guardian

Marsch is applying the same theories of togetherness and holism that he refined with New York Red Bulls while busting preconceptions about American managers abroad
Pre-season, 2016. New York Red Bulls head coach Jesse Marsch has arranged a training camp in Tucson, Arizona, for both his own squad and the club’s second team. Rather than separate the two groups and focus his attention of the first-teamers, Marsch brings everyone together, delivering a training session for 45 players of varied ages and experience levels.
On the sideline, Marsch pauses mid-session. Wearing a contented smile, he turns to a colleague. “This,” he says, “I love.”
Six years on, Marsch is applying the same theories of togetherness and holism in his role as Leeds United manager that he refined in New York, with stops in between – with mixed results – in Austria and Germany.
It will come as little surprise to those who have worked with him, then, to learn that, for example, at a recent press conference Marsch chose to discuss the exam results of youth-team midfielder Archie Gray, and how he’d encouraged the teen to pursue further education.
“He gets everyone on the same page – the staff, the players, the guys that are taking care of the pitches, the people cleaning the facilities,” says former Leeds winger Mike Grella, who played under Marsch with the Red Bulls. “Everyone involved with the club, he’ll spend time with them, get their take, get everyone fighting to win things together. That’s what he does best: he gets the best out of people.”
And, for the most part, it’s working.
Marsch was appointed in January to replace Marcelo Bielsa. Leeds, in just their second season back in the Premier League after a 16-year absence, were in danger of relegation. He immediately moved away from the man-marking style Bielsa had introduced and implemented a 4-2-2-2 shape used widely throughout the Red Bull clubs – which later evolved into a narrow 4-2-3-1 – and intense pressing off-ball.
Leeds improved enough to avoid the drop. This season, after a summer transfer window in which additions were made to suit his style, and with a full pre-season to work with the squad, Marsch will feel August’s 3-0 demolition of Chelsea at Elland Road was the full implementation of his style: relentless pressing and a compact, narrow attack.
Poor results since – a draw at home to Everton sandwiched between defeats away to Brighton and Brentford – evidence the remaining room for improvement. But if Marsch needed further buy-in – internally or externally – the Chelsea game was the perfect illustration of what is possible with his methods and Leeds’ carefully assembled collection of players.
Still, not everyone is convinced.
It isn’t just the spectre of an adored departing manager with which Marsch has had to contend but also the stigma against American coaches when they cross the Atlantic.
Upon accepting the Leeds job, Marsch sought the counsel of his former Princeton University and DC United coach Bob Bradley, the first American to manage in the Premier League. Bradley’s appointment at Swansea City was publicly condemned by the club’s supporters’ trust and he was routinely mocked for his use of American soccer vernacular. He was sacked after just 85 days.
Whatever wisdom he gleaned from Bradley has not prevented Marsch and his mannerisms from offending. His theatrical complaining to match officials in the 5-2 loss to Brentford, for example, earned him a red card and no shortage of criticism online and in print.
But the people Marsch most needed to win over, the Leeds players, have largely been impressed by the American. Those who played under Bielsa admired the idiosyncratic Argentinian, but he was more distant, less personable, than the new boss. Players have privately spoken of Marsch, his approach and tactical acumen, in complimentary tones.
“He’s so personable and he’s so positive as a man,” says David Longwell, a first-team coach with Shrewsbury Town who worked with Marsch when he was NYRB’s academy manager. “He’s just a leader. Every organization can talk about culture, but at Red Bulls the culture came from Jesse. He drove the culture.
“When you’re around him, you want to do well for him, you want to work hard for him. That’s players and staff.”
“He has a great sense of when to make it business and when to make it about personal relationships,” adds San Jose Earthquakes technical director John Wolyniec, who was previously the coach of New York Red Bulls II. “He was always excited to bring players in and, when most coaches would show clips [of games], he’d show pictures of their family.”
Leeds had a tenuous American connection before Marsch arrived: winger Jack Harrison, though born in Stoke-on-Trent, is a product of the US soccer system. He left Manchester United’s academy at 14 to attend Berkshire School in Massachusetts; he then spent two years with Manhattan SC and played college football for the Wake Forest Demon Deacons, before signing with Manchester City sister club New York City FC.
But the summer arrivals of Brenden Aaronson and Tyler Adams have served to bolster a transatlantic identity within the West Yorkshire side.
Marsch has worked with both players before: Aaronson at Red Bull Salzburg, from whom Leeds bought the 21-year-old for $28.8m (£24.7m); and Adams, who was signed from RB Leipzig for $23.1m (£20m), as a youngster in New York.
Aaronson impressed hugely in the Chelsea victory. His goal, for which he harried Blues keeper Edouard Mendy into a costly error, was indicative the attacking midfielder’s bustling, high-octane playing style.
And just as Marsch is tasked with filling the hefty void left by Bielsa, Adams has stepped into the deep-midfield role previously occupied by Leeds native Kalvin Phillips, who was sold to Manchester City this summer for $51.9m (£45m).
Those who have worked with Adams are sure the 23-year-old will not be overawed by the prospect of replacing Phillips, nor of the physical rigors of the Premier League.
“Toughness is probably his biggest thing,” Grella says of his former NYRB team-mate. “And it’s real toughness. You look at him and he’s not very big, he’s not very imposing. But the toughness is within – his attitude, his ferociousness, his ability to cover so much ground, his ability to be a good team-mate.
“He got into a little scrap with one of the senior players and held his own. He actually kind of beat him up a little bit. Even though he was just as kid, to draw a line and physically stand up for yourself, it was impressive. From that day, we were like, ‘Yeah, this kid’s legit.’
“He epitomizes everything Jesse wants to do: from the way he wants to play to his mentality, character, all that. Tyler Adams is the perfect fit for what Jesse wants.”
That Aaronson and Adams are thriving in tandem for Leeds in the Premier League is to the great advantage of the United States ahead of the World Cup this winter.
And while the American duo have proven they have the technical quality to succeed at the highest level, it is their intensity and commitment that has most impressed fans and Marsch alike.
In New York, the manager was renowned for arriving early at the club’s training ground, clocking in daily before 7am, and he even wheeled an exercise bike into his office so he could combine a physical workout with his technical planning. He has spoken of a desire to learn about Yorkshire and wanting his Leeds side to reflect the hard-working ethos of the county’s people.
“The one thing Leeds fans will always support and always fall in love with is your work ethic,” says Grella. “No one’s got a better work ethic than Jesse, Tyler and Aaronson. So as long as they keep that up they’ll always have the respect and the hearts of the Leeds fans.”
But Marsch know he will never please all the people all of the time. He is instead focused on continuing to mould Leeds and their new transatlantic core in his image.
“There is probably still a lot of doubt in me,” he said the day after his side hammered Chelsea at Elland Road. “It’s OK. It’s normal. There are going to be people that like me and people that hate me.
“I just want the team to play with love, passion and belief.”

source

‹ Previous Next ›