DeepGestalt

๋ถ„์•ผ
Image classification
๋ฆฌ๋ทฐ ๋‚ ์งœ
2020/07/27
๋ณธ ํฌ์ŠคํŠธ๋Š” ์ œ๊ฐ€ ํœด๋จผ์Šค์ผ€์ดํ”„ ๊ธฐ์ˆ  ๋ธ”๋กœ๊ทธ์— ๋จผ์ € ์ž‘์„ฑํ•˜๊ณ  ์˜ฎ๊ธด ํฌ์ŠคํŠธ์ž…๋‹ˆ๋‹ค.
๋ณธ ํฌ์ŠคํŠธ์—์„œ๋Š” ๋”ฅ๋Ÿฌ๋‹์œผ๋กœ ํฌ๊ท€ ์œ ์ „์„ฑ ์งˆํ™˜์„ ์ง„๋‹จํ•œ๋‹ค๋Š” ๋‚ด์šฉ์˜ ๋…ผ๋ฌธ์— ๋Œ€ํ•ด์„œ ๋ฆฌ๋ทฐํ•˜๋ ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๋…ผ๋ฌธ์˜ ์ œ๋ชฉ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค
โ€œDeepGestalt-Identifying Rare Genetic Syndromes Using Deep Learningโ€
๋…ผ๋ฌธ์— ๋Œ€ํ•œ ๋‚ด์šฉ์„ ์ง์ ‘ ๋ณด์‹œ๊ณ  ์‹ถ์œผ์‹  ๋ถ„์€ย ์ด๊ณณ์„ ์ฐธ๊ณ ํ•˜์‹œ๋ฉด ์ข‹์Šต๋‹ˆ๋‹ค.
DeepGestalt: High level flow

Objective

๋…ผ๋ฌธ์—์„œ ๋ชฉ์ ์œผ๋กœ ํ•˜๋Š” ๋ฐ”๋Š” ํ™˜์ž์˜ ์–ผ๊ตด ์‚ฌ์ง„์„ ์ด์šฉํ•ด ๊ทธ ํ™˜์ž์˜ ํฌ๊ท€ ์งˆํ™˜์„ ์˜ˆ์ธกํ•ด๋‚ด๋Š” ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์„ ๊ตฌํ˜„ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
๋…ผ๋ฌธ์—์„œ ์ธ์šฉํ•œ ์ž๋ฃŒ์— ๋”ฐ๋ฅด๋ฉด ๋งŽ์€ ์œ ์ „์„ฑ ์งˆํ™˜๋“ค์ดย ์–ผ๊ตด๋กœ ํ‘œํ˜„๋˜๋Š” ํ‘œํ˜„ํ˜•(facial phenotype)์„ ๊ฐ€์ง€๊ธฐ์— ๊ฐ€๋Šฅํ•œ ๋ฐฉ๋ฒ•์ด๋ฉฐ, ์–ผ๊ตด ์‚ฌ์ง„์€ ์œ ์ „ํ•™์ž๋“ค์ด ์œ ์ „๋ณ‘์„ ์ง„๋‹จํ•˜๋Š”๋ฐ ์‹ค์ œ๋กœ ๋„์›€์„ ์ฃผ๊ณ  ์žˆ๋Š” ๋ฐ์ดํ„ฐ๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค.
์ด๋Ÿฌํ•œ ๊ธฐ๋ฐ˜๊ณผ ๋ชฉ์ ์„ฑ์„ ๊ฐ€์ง€๊ณ  ๋…ผ๋ฌธ์—์„œ ๊ตฌ์ฒด์ ์œผ๋กœ ๊ตฌํ˜„ํ•˜๊ณ ์ž ํ•˜๋Š” blackbox model์€ ํ™˜์ž์˜ ์–ผ๊ตด ์‚ฌ์ง„ x๋ฅผ input์œผ๋กœ ๋„ฃ์–ด similarity score(๋…ผ๋ฌธ์—์„œ๋Š” ์ด๋ฅผ Gestalt score๋ผ๊ณ  ์ •์˜ํ•ฉ๋‹ˆ๋‹ค)๊ฐ€ ํฌํ•จ๋œ ์œ ์ „์„ฑ ์งˆํ™˜์˜ sorted list๋ฅผ output์œผ๋กœ ์‚ฐ์ถœํ•˜๋Š” ํ•จ์ˆ˜ f(x)์ž…๋‹ˆ๋‹ค.

Methods

Image preprocessing

๊ฐ€์žฅ ์ฒซ ๋‹จ๊ณ„๋Š” image preprocessing ๋‹จ๊ณ„์ž…๋‹ˆ๋‹ค. Image processing ๋‹จ๊ณ„์—์„œ ์ค‘์š”์‹œํ•˜๊ฒŒ ์ƒ๊ฐํ•œ ๊ฒƒ์€ input image ๋“ค์˜ alignment์ž…๋‹ˆ๋‹ค.
๊ทธ ์ค‘ ๋‹จ์—ฐ ์ฒซ ๋‹จ๊ณ„๋Š”ย ๋‚  ๊ฒƒ์˜ ์ด๋ฏธ์ง€๋กœ๋ถ€ํ„ฐ ์–ผ๊ตด์„ ์ธ์‹ํ•ด๋‚ด๋Š” ๋‹จ๊ณ„์ž…๋‹ˆ๋‹ค. ํ˜„์‹ค ์„ธ๊ณ„ ์† ์กฐ์ž‘๋˜์ง€ ์•Š์€ ์ด๋ฏธ์ง€๋ฅผ input ์œผ๋กœ ๋ฐ›์„ ์ˆ˜ ์žˆ๊ธฐ ์œ„ํ•ด์„œ ํ•„์š”ํ•œ ๋‹จ๊ณ„์ž…๋‹ˆ๋‹ค. ๋…ผ๋ฌธ์—์„œ๋Š” โ€œA convolutional neural network cascade for face detectionโ€ ์—์„œ ๊ตฌํ˜„ํ•œ Deep Convolutional Neural Network(DCNN) ์ด์–ด๋ถ™์ธ ๋ชจ๋ธ์˜ ์–ผ๊ตด ์ธ์‹์„ ์ด์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.
Test pipeline of facial detector [์ถœ์ฒ˜: A convolutional neural network cascade for face detection]
๋‹ค์Œ ๋‹จ๊ณ„๋Š”ย ์ œ๋ฉ‹๋Œ€๋กœ์ธ ๊ฐ๊ฐ์˜ ์ด๋ฏธ์ง€๋“ค์„ alignํ•˜๋Š” ๋‹จ๊ณ„์ž…๋‹ˆ๋‹ค. ์ด ๊ณผ์ •์€ ์ดํ›„ ์‚ฌ์šฉ๋  image combine์„ ๋น„๋กฏํ•ด input image์˜ fomat์„ ํ†ต์ผํ•˜์—ฌ ์ข‹
์€ ํ•™์Šตํšจ๊ณผ๋ฅผ ๊ฐ€์ ธ์˜ค๊ธฐ ์œ„ํ•œ ๋‹จ๊ณ„์ž…๋‹ˆ๋‹ค. ๋…ผ๋ฌธ์—์„œ๋Š” 130๊ฐœ์˜ ์–ผ๊ตด ์† ํŠน์ง•์ ์ธ ์ง€์ (facial landmarks)๋ฅผ ์ฐพ์•„๋‚ด์–ด ์ด๋ฅผ ์ด์šฉํ•ด alignment๋ฅผ ์ง„ํ–‰ํ•œ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. Alignment๋ฅผ ํ†ตํ•œ ์„ฑ๋Šฅ ๊ฐœ์„ ์€ ๋…ผ๋ฌธ โ€œLearning to Align from Scratchโ€์—์„œ ์ฐพ์•„๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Sample images from LFW produced by different alignment algorithms [์ถœ์ฒ˜: Learning to Align from Scratch]
๋งˆ์ง€๋ง‰ ๋‹จ๊ณ„๋Š”ย align๋œ ์ด๋ฏธ์ง€๋ฅผ ๊ณ ์ •๋œ ์‚ฌ์ด์ฆˆ 100 * 100์˜ ํ˜•ํƒœ๋กœ ๋ณ€ํ™˜ํ•˜๊ณ  ํšŒ์ƒ‰์กฐ(grayscale)๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ๊ณผ์ •์ž…๋‹ˆ๋‹ค. ๋”๋ถˆ์–ด ์–ผ๊ตด์˜ ํŠน์ • ์˜์—ญ(facial region)์— ๋Œ€ํ•œ ์„ ํƒ์ ์œผ๋กœ ์ด๋ฏธ์ง€๋ฅผ ์ž๋ฅด๋Š” ๊ณผ์ •๋„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ์ง„ํ–‰ํ•˜๋Š” ์ด์œ ๋Š” ๋ณธ ํฌ์ŠคํŠธ์˜ ๋งˆ์ง€๋ง‰์—์„œ ์„ค๋ช…๋“œ๋ฆฌ๊ฒ ์Šต๋‹ˆ๋‹ค. ์ด ๊ณผ์ •์„ ๋งˆ์น˜๋ฉด input์„ ๋ชจ๋ธ์— ๋„ฃ์„ ์ค€๋น„๊ฐ€ ์™„๋ฃŒ๋œ ๊ฒƒ์ž…๋‹ˆ๋‹ค.

Phenotype extraction & Syndromes classification

๋…ผ๋ฌธ์—์„œ ๊ฐ€์žฅ ๋„์ „์ ์ด์—ˆ๋˜ ๋ฌธ์ œ์  ์ค‘ ํ•˜๋‚˜๋Š” ํฌ๊ท€ ์งˆํ™˜์„ ๊ฐ€์ง„ ํ™˜์ž๋“ค์˜ ๋ฐ์ดํ„ฐ์˜ ์ˆ˜๊ฐ€ ํ•™์Šต์‹œํ‚ค๊ธฐ์— ๋ถ€์กฑํ–ˆ๋‹ค๋Š” ์ ์ด์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด์„œ ๋…ผ๋ฌธ์—์„œ๋Š” ๋‘ ๋‹จ๊ณ„๋กœ ํ•™์Šต์„ ์ง„ํ–‰์‹œ์ผฐ์Šต๋‹ˆ๋‹ค.
์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„๋Š”ย baseline face representation์„ ์œ„ํ•œ ํ•™์Šต์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์ด๋ฏธ์ง€ ์†์—์„œ ์–ผ๊ตด์„ ํ‘œํ˜„ํ•˜๋Š” ๋‹จ๊ณ„์ž…๋‹ˆ๋‹ค. ์ด์ „์— image preprocessing ์—์„œ๋„ ์ง„ํ–‰ํ•œ ๋‹จ๊ณ„์ด์ง€๋งŒ, ์ด ๋‹จ๊ณ„์—์„œ๋„ ์ง„ํ–‰ํ•˜๋Š” ๊ฒƒ์€ ๋‹ค์Œ ๋‹จ๊ณ„์—์„œ ์ง„ํ–‰ํ•  fine-tuning์˜ ์ดˆ๊ธฐ ๊ฐ’(initial weight)์„ ๋ชจ๋ธ์— ์„ค์ •ํ•ด๋‘๊ธฐ ์œ„ํ•จ์ž…๋‹ˆ๋‹ค.
๋‹ค์Œ ๋‹จ๊ณ„๋Š”ย genetic syndrome classification์„ ์œ„ํ•œ ํ•™์Šต์ž…๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„์—์„œ๋Š” ์ด์ „ ๋‹จ๊ณ„์—์„œ ๊ตฌํ•œ weight๊ฐ’์„ ์—…๋ฐ์ดํŠธํ•˜์—ฌ ์–ผ๊ตด ์ •๋ณด ํ‘œํ˜„์— ๊ทธ์ณค๋˜ ๋ชจ๋ธ์ด ์–ผ๊ตด ์ •๋ณด ํ‘œํ˜„๊ณผ ๋”๋ถˆ์–ด ๊ทธ ํ‘œํ˜„ ์† ์œ ์ „์„ฑ ์งˆํ™˜ ์ •๋ณด๋ฅผ ํ•™์Šตํ•˜๊ธฐ ์œ„ํ•จ์ž…๋‹ˆ๋‹ค.
์•„๋ž˜๋Š” ๋ชจ๋ธ์˜ ์•„ํ‚คํ…์ณ์— ๋Œ€ํ•œ ๊ทธ๋ฆผ์ž…๋‹ˆ๋‹ค.
The Deep Convolutional Neural Network architecture of DeepGestalt

Choosing dataset

๋…ผ๋ฌธ์—์„œ๋Š” ๋ชจ๋ธ์˜ ํ•™์Šต์— ํ•„์š”ํ•œ dataset์œผ๋กœ CASIA Web-Face dataset ๊ณผ Face2Gene phenotype dataset์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.
CASIA Web-Face dataset์€ baseline face representation์„ ํ•™์Šต์‹œํ‚ค๊ธฐ ์œ„ํ•œ ์šฉ๋„๋กœ ์‚ฌ์šฉ๋˜์—ˆ๊ณ  10575๋ช…์˜ ์‚ฌ๋žŒ์œผ๋กœ๋ถ€ํ„ฐ ์–ป์–ด์ง„ 494414 images ๋ฅผ ํฌํ•จํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
Face2Gene phenotype dataset์€ genetic syndrome classification์„ ํ•™์Šต์‹œํ‚ค๊ธฐ ์œ„ํ•œ ์šฉ๋„๋กœ ์‚ฌ์šฉ๋˜์—ˆ๊ณ  2500๊ฐœ์˜ ์งˆํ™˜์— ๋Œ€ํ•œ ์ˆ˜๋งŒ๊ฐœ์˜ images ๋ฅผ ํฌํ•จํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

Training

๋”ฅ๋Ÿฌ๋‹์—์„œ weight์˜ ์ดˆ๊ธฐ๊ฐ’ ์„ค์ •์€ ๊ต‰์žฅํžˆ ์ค‘์š”ํ•œ ์ผ์ž…๋‹ˆ๋‹ค. ์ดˆ๊ธฐ๊ฐ’ ์„ค์ •์— ๋”ฐ๋ผ ํ•™์Šต์ด ์˜ฌ๋ฐ”๋ฅธ ๋ฐฉํ–ฅ์œผ๋กœ ๊ฐˆ ์ˆ˜๋„ ์žˆ๊ณ , local minimum์— ๋น ์ ธ์„œ ํ•™์Šต์— ์‹คํŒจํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด์„œ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด weight initializer ์ž…๋‹ˆ๋‹ค.
์—ฌ๋Ÿฌ ๋ฒˆ์˜ ์‹œ๋„ ๋์— ๋…ผ๋ฌธ์—์„œ๋Š” baseline face representation ์—์„œ๋Š” He Norma Initializer๋ฅผ, genetic syndrome classification์—์„œ๋Š” Xavier normal initializer๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๊ฐ€์žฅ ์ข‹์€ ์„ฑ๋Šฅ์„ ๋‚ด๋Š” initializer๋ฅผ ์„ ํƒ์ ์œผ๋กœ ์ฑ„์šฉํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
๋˜ํ•œ, ๋…ผ๋ฌธ์—์„œ๋Š” augmentation ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ์˜ ๊ฐœ์ˆ˜๋ฅผ ๋Š˜๋ฆฌ๊ณ  ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋…ผ๋ฌธ์—์„œ ์กฐ์ž‘ํ•œ ๊ฒƒ์€ image์˜ ํšŒ์ „, vertical/horizontal shift, zoom, shear transformation ์ •๋„์ž…๋‹ˆ๋‹ค.
์ด ์™ธ์—๋„ ๋…ผ๋ฌธ์—์„œ ์–ธ๊ธ‰ํ•œ ๊ฒƒ์€ learning rate, epoch, momentum ๋“ฑ์˜ ์ •๋ณด์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์ถ”ํ›„์— deep learning์— ๋Œ€ํ•œ ๊ฐœ๊ด€์— ๋Œ€ํ•ด์„œ ์„ค๋ช…ํ•  ๋•Œ ๋ชจ์•„์„œ ์ง„ํ–‰ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.
์ง€๊ธˆ์€ ๋ฐ์ดํ„ฐ๋ฅผ ์ด์šฉํ•ด training์„ ํ•  ๋•Œ ๋ณธ ๋…ผ๋ฌธ์—์„œ ์ค‘์š”์‹œ ํ•œ ์ ์ดย weight initialization์ธ ๊ฒƒ๊ณผย augmentationย ์ด๋ผ๋Š” ์ ๋งŒ ์งš๊ณ  ๋„˜์–ด๊ฐ€๊ฒ ์Šต๋‹ˆ๋‹ค.

Evalutaition

๋…ผ๋ฌธ์—์„œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ์šฉํ•œ ๊ฒƒ์€ย top-K-accuracyย ์ž…๋‹ˆ๋‹ค.
Top-K-accuracy๋Š” sorted list ํ˜•ํƒœ๋กœ ๊ฐ€๋Šฅ์„ฑ์„ ์˜ˆ์ธกํ•œ ๋ชจ๋ธ์—์„œ K๋ฒˆ์งธ ์•ˆ์— ์‹ค์ œ ๊ฐ’์ด ์กด์žฌํ•˜๋Š” ๊ฒฝ์šฐ๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ์ฒ™๋„์ž…๋‹ˆ๋‹ค.
์ฆ‰, top-10-accuracy ์˜ ๊ฒฝ์šฐ, ์ „์ฒด test ๋ฐ์ดํ„ฐ๋“ค ์ค‘ ๊ฐ€๋Šฅ์„ฑ์„ ์˜ˆ์ธกํ•œ sorted list์˜ 10๋ฒˆ์งธ ์ˆœ์„œ ์•ˆ์— ์‹ค์ œ๊ฐ’์ด ์กด์žฌํ•  ๊ฒฝ์šฐ์˜ ์ˆ˜๋กœ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

Experiments and Results

๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ํฌ๊ฒŒ 3๊ฐ€์ง€์˜ ํ˜•ํƒœ๋กœ ์‹คํ—˜์„ ์ง„ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ์ค‘์—์„œ ๋…ผ๋ฌธ์—์„œ ๊ฐ€์žฅ ์ค‘์š”ํ•˜๊ฒŒ ์ƒ๊ฐํ–ˆ๋˜ multi-class Gestalt Model์— ๋Œ€ํ•ด์„œ๋งŒ ์–ธ๊ธ‰ํ•˜๋„๋ก ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.
(๋‚˜๋จธ์ง€ ๋‘ ๊ฐœ๋Š” ๊ฐ๊ฐ ๊ตฌํ˜„ํ•œ Gestalt Model์ด binary classfication problem์—๋„ ์ž˜ ์ ์šฉ๋˜๋Š”์ง€์™€ genotype ์˜ˆ์ธก์—๋„ ์ž˜ ์ ์šฉ๋˜๋Š”์ง€์— ๋Œ€ํ•œ ์‹คํ—˜์ด์—ˆ์Šต๋‹ˆ๋‹ค)
Multi-class Gestalt Model ์€ ๋…ผ๋ฌธ์—์„œ ๋ชฉํ‘œ๋กœ ํ–ˆ๋˜ ๋ฐ”๋ฅผ ๊ตฌํ˜„ํ•œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค.ย ๋‹ค์–‘ํ•œ ์งˆํ™˜์„ ๊ฐ€์ง„ ํ™˜์ž๋“ค์ด ์„ž์—ฌ ์žˆ๋Š” ์‚ฌ์ง„ ์†์—์„œ ๊ฐ๊ฐ์˜ ์‚ฌ์ง„์— ๋Œ€ํ•ด์„œ ์‚ฌ์ง„ ์† ์‚ฌ๋žŒ์ด ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ํฌ๊ท€ ์งˆํ™˜์˜ ๊ฐ€๋Šฅ์„ฑ์„ ํ‘œํ˜„ํ•œ sorted list๋ฅผ ์‚ฐ์ถœํ•˜๋Š” ๊ฒƒ์ด ์ด ๋ชจ๋ธ์˜ ์—ญํ• ์ด์—ˆ์Šต๋‹ˆ๋‹ค.
216๊ฐœ์˜ ์œ ์ „์„ฑ ํฌ๊ท€์งˆํ™˜์„ ๊ฐ€์ง„ 26190๊ฐœ์˜ image๋ฅผ ํ†ตํ•ด ํ•™์Šต๋˜๊ณ , 502๊ฐœ์˜ ์‹ค์ œ ํ™˜์ž์˜ image๋ฅผ ํ†ตํ•ด ํ…Œ์ŠคํŠธ๊ฐ€ ์ง„ํ–‰๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
DeepGestalt performance and permutation test result
์œ„ ๊ฒฐ๊ณผ๋Š” model์˜ accuracy ์™€ permutation test์˜ mean value ๋ฅผ ๋น„๊ตํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ๋•Œ permutation test ์˜ ๊ฒฝ์šฐ mean value๋ฅผ test set์œผ๋กœ ์‚ฌ์šฉํ•œ ๋ฐ์ดํ„ฐ์˜ label(์—ฌ๊ธฐ์„œ๋Š” ํ™˜์ž์˜ ์งˆํ™˜)์„ ๋งˆ์Œ๋Œ€๋กœ ์„ž์€ ํ›„ ๊ทธ ์ค‘ ๋™์ผํ•˜๊ฒŒ ๋งž์„ ๊ฒฝ์šฐ์˜ ์ˆ˜๋ฅผ ํ†ตํ•ด์„œ ์ธก์ •ํ•ฉ๋‹ˆ๋‹ค.
๋‹น์—ฐ์Šค๋Ÿฝ๊ฒŒ๋„, ์•„๋ฌด๊ฒƒ๋„ ํ•˜์ง€ ์•Š์€ ์ฑ„ ์ถ”์ธกํ•œ ๊ฒƒ๋ณด๋‹ค ๋…ผ๋ฌธ์—์„œ ์ œ์‹œํ•œ ๋ชจ๋ธ์„ ์ด์šฉํ•ด ํ•™์Šตํ•˜์—ฌ ์‚ฐ์ถœํ•œ ๊ฒฐ๊ณผ ๊ฐ’์ด ๋” ๋†’์€ ์ •ํ™•์„ฑ์„ ๊ฐ€์ง€๊ณ  ์˜๋ฏธ์žˆ๋Š” ๊ฐ’์„ ๋‚˜ํƒ€๋‚ด๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Reason why the paper used cropped images and combine them
์•ž์„œ ๋ง์”€๋“œ๋ฆฐ ๋‚ด์šฉ ์ค‘์— ์งš๊ณ  ๋„˜์–ด๊ฐ€์ง€ ์•Š์€ ๋ถ€๋ถ„์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋…ผ๋ฌธ์—์„œ image์˜ ๊ฐ๊ฐ์˜ ์˜์—ญ์„ ์„ ํƒ์ ์œผ๋กœ ์ž๋ฅด๋Š” ๊ณผ์ •์„ ์ง„ํ–‰ํ•œ๋‹ค๊ณ  ํ–ˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ถ€๋ถ„์— ๋Œ€ํ•ด์„œ ๋…ผ๋ฌธ์—์„œ ๊ฐ ์˜์—ญ ๋ณ„๋กœ์˜ ํ•™์Šต ๋ฐ์ดํ„ฐ๋กœ ์ถ”์ธกํ•œ accuracy ์™€ ์ด๋ฅผ ํ•ฉ์นœ ๋ชจ๋ธ๋กœ ์ถ”์ธกํ•œ accuracy๋ฅผ ๋น„๊ตํ•œ ๊ฒฐ๊ณผ๋ฅผ ์ œ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค.
ํ‘œ๋ฅผ ๋ณด์‹œ๋ฉด Full Face ๋กœ ์ถ”์ธกํ•œ ๊ฒƒ๋ณด๋‹ค Aggregated model ๋กœ ์ถ”์ธกํ•œ ๊ฒƒ์ด ๋” ๋†’์€ accuracy๋ฅผ ๊ฐ€์ง์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๊ฒƒ์ด ๋…ผ๋ฌธ์—์„œ input image ๋ฅผ ๋„ฃ์„ ๋•Œ ๊ตณ์ด ๊ณผ์ •์„ ๋”ํ•ด๊ฐ€๋ฉด์„œ image๋ฅผ ์ž๋ฅด๊ณ  ๊ฐ๊ฐ์„ ํ•™์Šต์‹œํ‚จ ์ด์œ ์ž…๋‹ˆ๋‹ค.

Conclusion

์ด๊ฒƒ์œผ๋กœ ๋…ผ๋ฌธย โ€œDeepGestalt-Identifying Rare Genetic Syndromes Using Deep Learningโ€์˜ ๋‚ด์šฉ์„ ๊ฐ„๋‹จํ•˜๊ฒŒ ์š”์•ฝํ•ด๋ณด์•˜์Šต๋‹ˆ๋‹ค. ์ค‘๊ฐ„์ค‘๊ฐ„์— ์–ธ๊ธ‰ํ•˜์ง€๋Š” ์•Š์•˜์ง€๋งŒ ๋…ผ๋ฌธ์—์„œ ๊ฐ•์กฐํ•œ ๋…ผ๋ฌธ์˜ ์˜์˜๋Š” randomํ•˜๊ฒŒ sampling ๋œ images ์— ๋Œ€ํ•ด์„œ ๋†’์€ ์ˆ˜์ค€์˜ accuracy ๋กœ ์œ ์ „์„ฑ ํฌ๊ท€ ์งˆํ™˜์„ ์ง„๋‹จํ•ด๋‚ผ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์ด์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ˜•ํƒœ์˜ ๋”ฅ๋Ÿฌ๋‹์ด ๋ฏธ๋ž˜ ์ •๋ฐ€ ์˜ํ•™์—์„œ ์ค‘์š”ํ•œ ์—ญํ• ์„ ์ฐจ์ง€ํ•  ๋‚ ์ด ๋จธ์ง€ ์•Š์„ ๊ฒƒ ๊ฐ™๋‹ค๊ณ  ๋Š๊ปด์ง‘๋‹ˆ๋‹ค.