๋ณธ๋ฌธ ๋ฐ”๋กœ๊ฐ€๊ธฐ
Deep Learning/[๋…ผ๋ฌธ] Paper Review

U-Net

by ์ œ๋ฃฝ 2023. 7. 5.
728x90
๋ฐ˜์‘ํ˜•

 

 

 

1. Intro
  • ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” CNN์˜ ์„ฑ๊ณต์ด Training Set์˜ ์–‘์ด ์ปค์ง€๋ฉด์„œ ์ƒ๊ธด ์ œํ•œ์ ์ธ ์ด์œ ๋ผ๊ณ  ๋งํ•จ.
  • ์ด์ „๊นŒ์ง€๋Š” CNN์€ Classification์„ ์œ„ํ•ด ๋งŽ์ด ์‚ฌ์šฉ๋˜์—ˆ์œผ๋‚˜ ์ƒ๋ฌผํ•™ ๋ถ„์•ผ์˜ ์˜์ƒ ์ฒ˜๋ฆฌ์—์„œ๋Š” Localization์ด ์ค‘์š”ํ–ˆ๊ณ , Semantic Segmentation์˜ ์ค‘์š”๋„๊ฐ€ ๋†’์•˜์Œ.
  • ํ•˜์ง€๋งŒ ์ƒ๋ฌผํ•™์— ๋Œ€ํ•œ Sample์˜ ๊ฐœ์ˆ˜๊ฐ€ 1000๊ฐœ๋ฐ–์— ๋˜์ง€ ์•Š๋Š” ๊ฒƒ์ด ๋‹ค์ˆ˜.
  • ๊ธฐ์กด์— ์‚ฌ์šฉํ•˜๋˜ sliding-window 2๊ฐ€์ง€ ๋‹จ์ 
  1. redundancy of over lapping patch(๊ฒน์น˜๋Š” ํŒจ์น˜์˜ ๋ถˆํ•„์š”ํ•œ ์ค‘๋ณต์„ฑ)์œ„์˜ ์‚ฌ์ง„์—์„œ ๋ณด์ด๋Š” ๊ฒƒ๊ณผ ๊ฐ™์ด patch๋ฅผ ์˜ฎ๊ธฐ๋ฉด์„œ ์ค‘๋ณต์ด ๋ฐœ์ƒํ•˜๊ฒŒ ๋จ=> ์ด ์ค‘๋ณต๋œ ๋ถ€๋ถ„์€ ์ด๋ฏธ ํ•™์Šต๋œ(๊ฒ€์ฆ๋œ) ๋ถ€๋ถ„์„ ๋‹ค์‹œ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์ด๋ฏ€๋กœ ๋˜‘๊ฐ™์€ ์ผ์„ ๋ฐ˜๋ณตํ•˜๋Š” ๊ฒƒ๊ณผ ๊ฐ™์Œ. ์ฆ‰, ๋ถˆํ•„์š”ํ•œ ์ค‘๋ณต์— ๋Œ€ํ•œ ๋‚ด์šฉ๋„ ํ•™์Šตํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์†๋„๋„ ๋Š๋ฆฌ๊ณ  ์‹œ๊ฐ„๋„ ์˜ค๋ž˜ ๊ฑธ๋ฆผ
  1. trade-off between localization accuracy and use of context patch ์‚ฌ์ด์ฆˆ๊ฐ€ ํฌ๋ฉด, max pooling์ด ๋” ๋งŽ์ด ์ ์šฉ ๋˜๊ณ  ์ •ํ™•ํ•œ ์œ„์น˜ ์ •๋ณด๋ฅผ ์•Œ๊ธฐ์—๋Š” ์–ด๋ ต์ง€๋งŒ, ๋” ๋„“์€ ๋ฒ”์œ„์˜ ์ด๋ฏธ์ง€๋ฅผ ๋ณด๊ธฐ ๋•Œ๋ฌธ์— context ์ธ์‹์—๋Š” ํšจ๊ณผ๋ฅผ ๊ฐ€์ง.

⇒ ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” Fully Convolutional network๋ฅผ ์†Œ๊ฐœํ•จ.

1.5 U-net: Improved Sliding Window Search Method - input
  • ๊ฒ€์ฆ์ด ๋๋‚œ ๋ถ€๋ถ„์€ ํ•˜์ง€ ์•Š๊ณ  ๋‹ค์Œ ํŒจ์น˜๋ถ€ํ„ฐ ์—ฐ์‚ฐ ์ง„ํ–‰

⇒ ๊ธฐ์กด์˜ sliding window ๋‹จ์  ํ•ด๊ฒฐ (์—ฐ์‚ฐ + ์†๋„ ๋ถ€๋ถ„)

  • ์ž…๋ ฅ์œผ๋กœ ์‚ฌ์šฉํ•จ
1.5 U-net: Overlap tile Method (Strategy) - input
  • U-net์˜ ๊ฒฝ์šฐ padding์„ ์‚ฌ์šฉํ•˜์ง€ ์•Š์Œ
  • ๋”ฐ๋ผ์„œ ์ถœ๋ ฅ ์ด๋ฏธ์ง€์˜ ํ•ด์ƒ๋„๊ฐ€ ์ž…๋ ฅ ์ด๋ฏธ์ง€๋ณด๋‹ค ํ•ญ์ƒ ์ž‘๊ธฐ ๋•Œ๋ฌธ์—, input image ํฌ๊ธฐ๋ฅผ ๋Š˜๋ ค์„œ ์‚ฌ์šฉ
  • ์˜ˆ๋ฅผ ๋“ค์–ด, ๋…ธ๋ž€์ƒ‰ ๋ถ€๋ถ„ ์˜์—ญ์˜ segmentation์ด ํ•„์š”ํ•˜๋ฉด ๊ทธ๊ฒƒ๋ณด๋‹ค ๋” ํฐ ๋ฒ”์œ„(ํŒŒ๋ž€์ƒ‰ ๋ฒ”์œ„)์˜ ํŒจ์น˜๋ฅผ ์ž…๋ ฅ์œผ๋กœ ๋„ฃ์–ด์„œ ๊ทธ ๊ฒฐ๊ณผ๋ฅผ ๋ฝ‘๊ณ , ์ถœ๋ ฅ ํŒจ์น˜๋กœ ์‚ฌ์šฉ.
  • ์ด๋ฏธ์ง€ ๊ฒฝ๊ณ„๋ถ€๋ถ„(์—†๋Š” ๋ถ€๋ถ„ ex) padding์—์„œ zero padding ๋ถ€๋ถ„)์€ ๋ฏธ๋Ÿฌ๋ง์„ ํ™œ์šฉ
  • ์œ„์˜ ์‚ฌ์ง„๊ณผ ๊ฐ™์ด ๊ฒน์น˜๋Š” ๋ถ€๋ถ„์ด ์ผ๋ถ€ ์กด์žฌํ•˜๊ฒŒ ๋จ
2. Network Architecture
  • ์ „์ฒด ๋„คํŠธ์›Œํฌ: 23 conv

 

  1. Contraction Path
    • CNN ์ด๋ฏธ์ง€์˜ context๋ฅผ ํฌ์ฐฉํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ด์คŒ.
  1. Expansive path
    • ์ž‘์•„์ง„ feature map์„ Upsampling ํ•ด์„œ ์›๋ณธ ์ด๋ฏธ์ง€์™€ ๋น„์Šทํ•œ ํฌ๊ธฐ๋กœ ๋Š˜๋ ค์ค€ ํ›„, Contracting Path์˜ feature map๊ณผ ๊ฒฐํ•ฉํ•˜์—ฌ(ํšŒ์ƒ‰ ํ™”์‚ดํ‘œ ๋ถ€๋ถ„) ๋” ์ •ํ™•ํ•œ ์œ„์น˜ ์ •๋ณด๋ฅผ ๊ฐ€์ง„ segmentation map์„ ์–ป๊ฒŒ ๋จ.

  1. Contraction path
    • ํŒŒ๋ž€์ƒ‰ ๋ถ€๋ถ„: conv
      • ๋‘๋ฒˆ์˜ 3x3 conv
      • ReLU ์‚ฌ์šฉ
    • ๋นจ๊ฐ„์ƒ‰ ๋ถ€๋ถ„: max pooling
      • 2x2 max pooling ์‚ฌ์šฉ
      • channel 2๋ฐฐ์”ฉ ๋Š˜๋ ค์คŒ
    1. Expansive Path
    • ์ดˆ๋ก์ƒ‰ ๋ถ€๋ถ„: up-sampling
      • ํฌ๊ธฐ๋ฅผ ํ‚ค์›Œ์คŒ (2๋ฐฐ์”ฉ)
    • ํŒŒ๋ž€์ƒ‰ ๋ถ€๋ถ„: conv
      • ๋‘๋ฒˆ์˜ 3x3 conv
      • ReLU ์‚ฌ์šฉ
    • ํšŒ์ƒ‰ ๋ถ€๋ถ„: concat
      • contraction path์—์„œ ์ถ”์ถœ๋œ feature map (๊ฒฝ๊ณ„ ๋ถ€๋ถ„์„ crop) ์„ concat
      • ์ขŒ์šฐ๋ฐ˜์ „ → ํ™•์žฅ
      • conv ์—ฐ์‚ฐ ์ˆ˜ํ–‰ํ•  ๋•Œ, ์ž…๋ ฅ ์ด๋ฏธ์ง€์˜ ๊ฒฝ๊ณ„ ๋ถ€๋ถ„์€ ์ปค๋„์ด ๊ฒน์ณ์ง€์ง€ ์•Š๊ธฐ์— ์ถœ๋ ฅ ์ด๋ฏธ์ง€์—์„œ ์†์‹ค๋  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธ
      • ์ฆ‰, ๋ณดํ†ต 3x3, 5x5 7x7 ํ•„ํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๋Š”๋ฐ, ๊ฐ€์žฅ์ž๋ฆฌ ๋ถ€๋ถ„์€ ๊ฒน์น˜๋Š” ๋ถ€๋ถ„์ด ๊ฑฐ์˜ ์—†์–ด์„œ ์ •๋ณด๊ฐ€ ์†์‹ค๋œ๋‹ค๊ณ  ํ‘œํ˜„ํ•จ
    • ์ฒญ๋ก์ƒ‰ ๋ถ€๋ถ„: 1x1 conv
      • ๋งˆ์ง€๋ง‰์€ class๋ฅผ 2๋กœ ์„ค์ • (๋ฐฐ๊ฒฝvs์„ธํฌ)

    โ€ป ์ถ”๊ฐ€ ๋…ผ๋ฌธ ๋‚ด์šฉ)
    1. ์—ฌ๊ธฐ์„œ x์™€ y๋Š” ์ž…๋ ฅ ์ด๋ฏธ์ง€์˜ ๊ฐ€๋กœ์™€ ์„ธ๋กœ ํฌ๊ธฐ๋ฅผ ์˜๋ฏธํ•จ.
    1. ์ตœ๋Œ€ ํ’€๋ง ์—ฐ์‚ฐ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์ž…๋ ฅ ์ด๋ฏธ์ง€ ํฌ๊ธฐ์˜ ์ ˆ๋ฐ˜์œผ๋กœ ์ค„์–ด๋“œ๋Š”๋ฐ, ์ด ๋•Œ ์ž…๋ ฅ ์ด๋ฏธ์ง€์˜ ํฌ๊ธฐ๊ฐ€ ํ™€์ˆ˜์ผ ๊ฒฝ์šฐ ์ •ํ™•ํžˆ ๋ฐ˜์œผ๋กœ ๋‚˜๋ˆ„๊ธฐ๊ฐ€ ์–ด๋ ค์›Œ์ง.
    1. ๋”ฐ๋ผ์„œ ์ž…๋ ฅ ์ด๋ฏธ์ง€์˜ ๊ฐ€๋กœ์™€ ์„ธ๋กœ ํฌ๊ธฐ๊ฐ€ ์ง์ˆ˜์ธ ๊ฒฝ์šฐ, ์ตœ๋Œ€ ํ’€๋ง ์—ฐ์‚ฐ์ด ๋ฐ˜์œผ๋กœ ์ค„์–ด๋“ค์—ˆ์„ ๋•Œ๋„ ๋ชจ๋“  ๋ ˆ์ด์–ด์—์„œ ์ ์šฉ๋˜๋„๋ก ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•จ.
3. Training
  1. ํ˜„๋ฏธ๊ฒฝ์œผ๋กœ ์ฐ์€ ์‚ฌ
  1. color์„ ๋‹ค๋ฅด๊ฒŒ ํ•œ (์ •๋‹ต ๋‹ต์ง€- ground truth)
  1. segmentation์„ black and white๋กœ ๋งŒ๋“  ๊ฒฐ๊ณผ๊ฐ’
  1. ์„ธํฌ ๊ฒฝ๊ณ„์„  ํ•™์Šต์‹œํ‚จ ์ด๋ฏธ์ง€

  • ๊ฒฝ์‚ฌํ•˜๊ฐ•๋ฒ• ์‚ฌ์šฉ
  • GPU ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์ตœ๋Œ€ํ•œ ํ™œ์šฉํ•˜๊ธฐ ์œ„ํ•ด ๋ฐฐ์น˜ ์‚ฌ์ด์ฆˆ๋ฅผ ์ค„์ด๊ณ , ํŒจ์น˜๋ฅผ ํฌ๊ฒŒ ํ–ˆ์Œ
  • ํ•˜์ง€๋งŒ, ๋ฐฐ์น˜ ์‚ฌ์ด์ฆˆ๊ฐ€ ์ž‘์€ ๊ฒฝ์šฐ, ์ตœ์ ํ™” ์ž˜ ์•ˆ๋จ
  • ์ด๋ฅผ ๋ณด์™„ํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋ฉ˜ํ…€์„ 0.99๋กœ → ๊ณผ๊ฑฐ์˜ ๊ฐ’์ด ๋งŽ์ด ๋ฐ˜์˜๋˜๋„๋ก.
  1. softmax
  • ์ „์ฒด ํด๋ž˜์Šค ์ค‘ ํ•ด๋‹น ํด๋ž˜์Šค์ผ ํ™•๋ฅ ๊ฐ’
  • ์—๋„ˆ์ง€ ํ•จ์ˆ˜: ์ตœ์ข… ํ”ผ์ณ ๋งต์— ๋Œ€ํ•œ ํ”ฝ์…€ ๋‹จ์œ„์˜ ์†Œํ”„ํŠธ๋งฅ์Šค์™€ ๊ต์ฐจ ์—”ํŠธ๋กœํ”ผ ์†์‹ค ํ•จ์ˆ˜์˜ ๊ฒฐํ•ฉ์œผ๋กœ ๊ณ„์‚ฐ๋จ.
  • ์„ธํฌ ์‚ฌ์ด์— ๊ฐ„๊ฒฉ์ด ์งง์•„์„œ ์„ธํฌ๋ณ„๋กœ ๊ตฌ๋ณ„์ด ํž˜๋“  ๊ฒฝ์šฐ๊ฐ€ ์žˆ๊ธฐ ๋•Œ๋ฌธ์—, ๊ฑฐ๋ฆฌ๊ฐ€ ๊ฐ€๊นŒ์šธ์ˆ˜๋ก ๊ฐ€์ค‘์น˜๋ฅผ ํฌ๊ฒŒ ํ•ด์„œ ๋ถ„๋ฆฌ๋ฅผ ํ™•์‹คํ•˜๊ฒŒ ํ•ด๋ฒ„๋ฆผ.
  • ์„ธํฌ ์‚ฌ์ด์— ๋–จ์–ด์ง„ ๊ฐ„๊ฒฉ์ด ์งง์•„ ์„ธํฌ๋ณ„๋กœ ๊ตฌ๋ณ„์ด ํž˜๋“  ๊ฒฝ์šฐ๊ฐ€ ๋ฐœ์ƒํ•˜๊ธฐ ๋•Œ๋ฌธ.

  • wc: ๊ฐ๊ฐ ํด๋ž˜์Šค๋งˆ๋‹ค ๋“ฑ์žฅํ•˜๋Š” ๋นˆ๋„์ˆ˜ ์กฐ์œจ( ex: ๋ฐฐ๊ฒฝ๊ณผ ์„ธํฌ์˜ ๋“ฑ์žฅ ํšŸ์ˆ˜ )
  • d1(x): ์ฒซ๋ฒˆ์งธ๋กœ ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด ์„ธํฌ๊นŒ์ง€์˜ ๊ฑฐ๋ฆฌ → ์—ฌ๊ธฐ์„œ x๋Š” ๋‘ ์„ธํฌ ์‚ฌ์ด์— ์กด์žฌํ•˜๋Š” ์ขŒํ‘œ๊ฐ’
  • d1(x): ๋‘ ๋ฒˆ์งธ๋กœ ๊ฐ€๊นŒ์šด ์„ธํฌ๊นŒ์ง€์˜ ๊ฑฐ๋ฆฌ
  • σ=5, w_0 =10

++ ๊ฑฐ๋ฆฌ๊ฐ€ ์งง์„์ˆ˜๋ก ๊ฐ€์ค‘์น˜๋ฅผ ํฌ๊ฒŒ

ex) d1=2, d2=4์ธ ๊ฒฝ์šฐ, ๊ฐ’์€ 0.00000152299

d1=1, d2=3์ธ ๊ฒฝ์šฐ, ๊ฐ’์€ 0.00335๋กœ ๊ฑฐ๋ฆฌ๊ฐ€ ์งง์„์ˆ˜๋ก ๊ฐ€์ค‘์น˜๊ฐ€ ํฌ๊ฒŒ ๋จ

์ฆ‰, ์ด ๋ง์€ ์„ธํฌ ๊ฐ„์˜ ๋ถ„๋ฆฌ๋ฅผ ํ™•์‹คํ•˜๊ฒŒ ํ•˜๊ฒ ๋‹ค!์˜ ์˜๋ฏธ๋ฅผ ๊ฐ€์งˆ ์ˆ˜ ์žˆ์Œ (๋ถ„๋ฆฌ๋ฅผ ๋” ์ž˜ํ•˜๋„๋ก)


  • ๊ฐ€์ค‘์น˜ ์ดˆ๊ธฐํ™”
    • ํ˜„ ๋‰ด๋Ÿฐ์— ๋“ค์–ด์˜ค๋Š” ๋…ธ๋“œ ๊ฐœ์ˆ˜๋ฅผ n์ด๋ผ๊ณ  ํ•œ๋‹ค๋ฉด, root(2/n)์˜ ํ‘œ์ค€ ํŽธ์ฐจ๋ฅผ ๊ฐ€์ง„ ๊ฐ€์šฐ์‹œ์•ˆ ๋ถ„ํฌ๋ฅผ ์ด์šฉํ•ด ๊ฐ€์ค‘์น˜๋ฅผ ์ดˆ๊ธฐํ™”
    • ReLU์™€ ์ž์ฃผ ์“ฐ์ด๋Š” He initialization ์‚ฌ์šฉํ•จ
    • ex) 3x3 CNN์— channel = 64์ธ feature map์ด ๋“ค์–ด์˜ค๋ฉด ํ•ด๋‹น CNN์˜ N = 9*64 = 576๊ฐœ์˜ ๊ฐ€์ค‘์น˜ ์ดˆ๊ธฐํ™”
3.1 (Training) Data augmentation
  • data augmentation์€ ํ•™์Šต ๋ฐ์ดํ„ฐ์…‹์ด ๋งŽ์ด ์—†์„ ๋•Œ ์œ ์šฉํ•จ
  • ํ˜„๋ฏธ๊ฒฝ ๋“ฑ์œผ๋กœ ์ดฌ์˜ํ•˜๋Š” ์‚ฌ์ง„๋“ค(microscopical image)์€ ์ƒ‰๊น”์ด ๋‹ค์–‘ํ•˜์ง€ ์•Š๊ณ  ํšŒ์ƒ‰๋น›๊น”๋กœ ์ด๋ฃจ์–ด์ ธ์žˆ๊ณ  ๊ฐ์ฒด๊ฐ„ ๊ตฌ๋ณ„๋„ ์„ ๋ช…ํ•˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์— Data Augmentation์„ ์ด์šฉํ•ด ํ’๋ถ€ํ•œ ๋ฐ์ดํ„ฐ์…‹์„ ๋งŒ๋“œ๋Š”๊ฒŒ ๋”์šฑ ํ•„์š”ํ•จ.
  • ์ผ๋ฐ˜์ ์ธ augmentation(์„ ํ˜•๋ณ€ํ™˜) + ์ถ”๊ฐ€์ ์œผ๋กœ Elastic Deformation ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉ → ๋น„์„ ํ˜•์ ์œผ๋กœ ๊ฐ€ํ•จ
  • ์„ธํฌ๊ฐ€ ์‚ด์•„์žˆ๊ธฐ์—, ์„ธํฌ๋Š” ํ•ญ์ƒ ๋™์ผ ๋ชจ์–‘์ด ์•„๋‹˜ → ์ˆœ๊ฐ„์˜ ๋ณ€ํ˜•๋“ค์„ ์ž˜ ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ .
4. Experiments

4.1 EM segmentation challenge

  • ์šฐ์„  U-Net์€ ์ „์žํ˜„๋ฏธ๊ฒฝ์œผ๋กœ ๊ด€์ฐฐ๋˜๋Š” ๋‰ด๋Ÿฐ ๊ตฌ์กฐ์—์„œ cell segmentation task๋ฅผ ์ˆ˜ํ–‰.

 - EM segmentation challenge์—์„œ ์ œ๊ณต๋˜๋Š” ๋ฐ์ดํ„ฐ์…‹์œผ๋กœ ํ•™์Šต.

  • ๋ฐ์ดํ„ฐ์…‹์€ ์ „์ž ํ˜„๋ฏธ๊ฒฝ์œผ๋กœ ์ฐ์€ 512 x 512 ํ•ด์ƒ๋„์˜ ์ด๋ฏธ์ง€ 30์žฅ์œผ๋กœ ์ด๋ฃจ์–ด์ ธ ์žˆ๊ณ  ์ด๋ฏธ์ง€์˜ ๊ฐ ๋ถ€๋ถ„์— ์„ธํฌ๋Š” ํฐ์ƒ‰์œผ๋กœ, ์„ธํฌ๋ง‰(membrane)์€ ๊ฒ€์€์ƒ‰์œผ๋กœ ์ƒ‰์น ํ•œ ground truth segmentation map์„ ๋งŒ๋“ฌ.
  • ํ…Œ์ŠคํŠธ์šฉ ์ด๋ฏธ์ง€๋„ ์žˆ๋Š”๋ฐ ground truth segmentation map์€ ๊ณต๊ฐœ๋˜์ง€ ์•Š์•˜๋‹ค๊ณ .
  • U-net์ด ๊ฐ€์žฅ ์ข‹์•˜๋‹ค!

4.2 ISBI cell tracking challenge

  • a, c๊ฐ€ ์ž…๋ ฅ ์ด๋ฏธ์ง€๊ณ  b, d๊ฐ€ ground truth segmentation map
  • a, c์™€ ๊ฐ™์ด ๊ด‘ํ•™ ํ˜„๋ฏธ๊ฒฝ์—์„œ ์–ป์€ ์ด๋ฏธ์ง€๋กœ b, d์™€ ๊ฐ™์ด ์„ธํฌ๋ฅผ ๊ตฌ๋ณ„ํ•˜๋Š” task๋ฅผ U-Net์ด ์–ผ๋งˆ๋‚˜ ์ž˜ ์ˆ˜ํ–‰ํ•˜๋Š”์ง€ ์‹œํ—˜ํ•ด๋ด„.
  • U-net์ด ์ข‹์•˜๋‹ค!
5. Conclusion
  • U-Net์€ ๋‹ค์–‘ํ•œ biomedical segmentation applications์—์„œ "์•„์ฃผ" ์ข‹์€ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์คŒ
  • ์ €์ž๋Š” elastic deformation์ด ํฌํ•จ๋œ Data augmentation ๋•๋ถ„์— ์ ์€ ์‚ฌ์ด์ฆˆ์˜ ๋ฐ์ดํ„ฐ์…‹๋งŒ ์š”๊ตฌํ–ˆ๊ณ  ํ•ฉ๋ฆฌ์ ์ธ ํ•™์Šต ์‹œ๊ฐ„(NVidia Titan GPU (6 GB)์—์„œ 10์‹œ๊ฐ„ ํ•™์Šต)์„ ๊ฐ€์กŒ๋‹ค๊ณ  ๋งํ•จ.
  • ๊ทธ๋ฆฌ๊ณ  ๋งˆ์ง€๋ง‰์œผ๋กœ U-Net์˜ ๊ตฌ์กฐ๊ฐ€ ๋‹ค์–‘ํ•œ task์— ์‰ฝ๊ฒŒ ์‘์šฉ๋  ์ˆ˜ ์žˆ์„๊ฑฐ๋ผ ํ™•์‹ ํ•œ๋‹ค๊ณ  ๋งํ•˜๋ฉฐ ๋…ผ๋ฌธ์„ ๋๋งˆ์นจ.
6. Reference

728x90
๋ฐ˜์‘ํ˜•

'Deep Learning > [๋…ผ๋ฌธ] Paper Review' ์นดํ…Œ๊ณ ๋ฆฌ์˜ ๋‹ค๋ฅธ ๊ธ€

Inception-v4, Inception-ResNetand the Impact of Residual Connections on Learning  (0) 2023.07.05
Seq2Seq  (0) 2023.07.05
Bert  (0) 2023.07.05
VIT [AN IMAGE IS WORTH 16X16 WORDS: TRANSFORMERS FOR IMAGE RECOGNITION AT SCALE]  (0) 2023.07.05
RetinaNet  (0) 2023.07.05