DRLN [Kor]

(Description) Saeed Anwar, Nick Barnes / Densely Residual Laplacian Super-Resolution / IEEE 2019

1. Problem definition

์ƒ๋Œ€์ ์œผ๋กœ ๋‚ฎ์€ ํ•ด์ƒ๋„์˜ ์ด๋ฏธ์ง€๋ฅผ ๋ณด๋‹ค ๋†’์€ ํ•ด์ƒ๋„๋กœ ๋ณต์›ํ•˜๋Š” ์ž‘์—…์„ ์ดˆํ•ด์ƒํ™”(Super-Resolution)๋ผ๊ณ  ํ•œ๋‹ค. ์ตœ๊ทผ ์ˆ˜๋…„๊ฐ„, ์ดˆํ•ด์ƒํ™” ์ž‘์—…์€ ๊ณ ํ•ด์ƒ๋„ ์ด๋ฏธ์ง€๋ฅผ ์š”ํ•˜๋Š” ์ž‘์—…๋“ค์— ์˜ํ•ด ์—ฐ๊ตฌ ์ˆ˜์š”๊ฐ€ ์ฆ๊ฐ€ํ•˜์˜€๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ๋‹จ์ผ ์ €ํ•ด์ƒ๋„ ์ด๋ฏธ์ง€๋ฅผ ์ดˆํ•ด์ƒํ™”ํ•˜๋Š” ์ž‘์—…์ธ Single Image Super-Resolution (SISR)์„ ๋ชฉํ‘œ๋กœ ํ•œ๋‹ค. ์ด๋ฏธ์ง€ ์ดˆํ•ด์ƒํ™”๋Š” ์ž…๋ ฅ๋˜๋Š” ์ €ํ•ด์ƒ๋„ ์ด๋ฏธ์ง€์— ๋Œ€์‘ํ•˜๋Š” ๊ณ ํ•ด์ƒ๋„ ์ด๋ฏธ์ง€ ์ถœ๋ ฅ์˜ ํฌ๊ธฐ๊ฐ€ ๋‹ฌ๋ผ์„œ 1๊ฐœ์˜ ์œ ์ผํ•œ ํ•ด๊ฐ€ ์กด์žฌํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹Œ ์—ฌ๋Ÿฌ ํ•ด๊ฐ€ ์กด์žฌํ•˜๊ฒŒ ๋˜๋Š” ๋ถˆ๋Ÿ‰์กฐ๊ฑด๋ฌธ์ œ(ill-posed problem)๊ฐ€ ๋ฐœ์ƒํ•œ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์‹ฌ์ธต ์ปจ๋ณผ๋ฃจ์…˜ ์‹ ๊ฒฝ๋ง(Deep Convolutional Neural Network,Deep CNN)์ด ์ ์šฉ๋˜์—ˆ๊ณ  ํ˜„์žฌ๊นŒ์ง€ ๋งŽ์€ ์ข…๋ฅ˜์˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ์—ฐ๊ตฌ๋˜์–ด ์™”๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ํ˜„์กดํ•˜๋Š” ์ดˆํ•ด์ƒํ™” ๋”ฅ๋Ÿฌ๋‹ ์•Œ๊ณ ๋ฆฌ์ฆ˜๋ณด๋‹ค ๋” ์ •ํ™•ํ•˜๊ณ  ์‹คํ–‰ ์‹œ๊ฐ„์ด ๋น ๋ฅธ ๋ชจ๋ธ์„ ์—ฐ๊ตฌํ•˜์˜€๋‹ค.

2. Motivation

ํ˜„์กดํ•˜๋Š” ์ดˆํ•ด์ƒํ™”๋ฅผ ์œ„ํ•œ Deep CNN ์•Œ๊ณ ๋ฆฌ์ฆ˜(SRCNN, RCAN ๋“ฑ)์€ ๋งค์šฐ ๋ณต์žกํ•œ ๊ตฌ์กฐ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์œผ๋ฉฐ, ๋ณต์žกํ•œ(Deep) ๋„คํŠธ์›Œํฌ์ผ ์ˆ˜๋ก ๊ธด ์‹คํ–‰์‹œ๊ฐ„์˜ ๋น„ํšจ์œจ์ ์ธ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ค€๋‹ค. ์ด์— ๋”ฐ๋ผ ๋„คํŠธ์›Œํฌ์˜ ๊นŠ์ด๋ฅผ ์ค„์—ฌ ํšจ์œจ์„ฑ์„ ๋†’์ธ ๋ชจ๋ธ๋“ค(DRCN, DRRN ๋“ฑ)์ด ์—ฐ๊ตฌ๋˜์—ˆ๋‹ค. ํ•˜์ง€๋งŒ ์ด๋Ÿฐ ๋ชจ๋ธ๋“ค์€ ์ด parameter์ˆ˜๋Š” ๊ฐ์†Œํ•˜๋”๋ผ๋„ ์ด ์—ฐ์‚ฐ๋Ÿ‰์€ ์ฆ๊ฐ€ํ•˜๊ฒŒ๋˜๋Š” ๋ฌธ์ œ๋ฅผ ๊ฐ–๊ณ ์žˆ๋‹ค. ๋’ค์ด์–ด ์ปจ๋ณผ๋ฃจ์…˜ ๊ณ„์ธต ๊ฐ„์˜ denseํ•œ ์—ฐ๊ฒฐ์„ ์ด์šฉํ•œ SRDenseNet๊ณผ RDN, parameter ์ˆ˜์™€ ์—ฐ์‚ฐ์†๋„๋ฅผ ๋ชจ๋‘ ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•ด group ์ปจ๋ณผ๋ฃจ์…˜์„ ์‚ฌ์šฉํ•œ CARN์ด ๋“ฑ์žฅํ–ˆ์œผ๋‚˜, ๋Œ€๋ถ€๋ถ„์˜ CNN ๋ชจ๋ธ์€ ํ•˜๋‚˜์˜ ์Šค์ผ€์ผ์„ ์‚ฌ์šฉํ•˜๊ฑฐ๋‚˜ ์—ฌ๋Ÿฌ ์Šค์ผ€์ผ์„ ์‚ฌ์šฉํ•˜๋”๋ผ๋„ ๊ฐ ์Šค์ผ€์ผ์˜ ๊ฐ€์ค‘์น˜๋ฅผ ๋™์ผํ•˜๊ฒŒ ๋ถ€์—ฌํ•˜๊ธฐ์— ๋‹ค์–‘ํ•œ ํ•ด์ƒ๋„์— ๋”ฐ๋ฅธ ์ ์‘๋ ฅ์ด ๋–จ์–ด์ง„๋‹ค.

Idea

  1. ์ดˆํ•ด์ƒํ™”์˜ ์ •ํ™•๋„ ํ–ฅ์ƒ์„ ์œ„ํ•ด ์ €ํ•ด์ƒ๋„์˜ ์ •๋ณด๋ฅผ ์ถฉ๋ถ„ํžˆ ์ด์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ ์šฉํ•˜์˜€๋‹ค.

  2. Densely connected residual block์—์„œ๋Š” ์—ฌ๋Ÿฌ๋ฒˆ shortcut์„ ์‚ฌ์šฉํ•˜์—ฌ ์›๋ž˜ ์ด๋ฏธ์ง€์˜ ์ •๋ณด๋ฅผ ํฌํ•จํ•œ feature๋ฅผ ๋™์‹œ์— ํ•™์Šตํ•œ๋‹ค.

  3. Laplacian attention network๋ฅผ ํ†ตํ•ด ์—ฌ๋Ÿฌ ์Šค์ผ€์ผ์˜ feature ์ •๋ณด๋ฅผ ํ•™์Šตํ•˜๋ฉฐ, ๋ชจ๋ธ๊ณผ feature ์‚ฌ์ด์˜ ์˜์กด๋„๋ฅผ ํ•™์Šตํ•œ๋‹ค.

3. Method

๋„คํŠธ์›Œํฌ ๊ตฌ์กฐ

์‚ฌ์šฉ๋œ ๋„คํŠธ์›Œํฌ๋Š” ํฌ๊ฒŒ 4๊ฐœ์˜ ๊ตฌ์กฐ(Feature ์ถ”์ถœ, Residual ๊ตฌ์กฐ ์—ฐ์‡„ ์ง„ํ–‰, Upsampling, ์ด๋ฏธ์ง€ ์žฌ๊ตฌ์„ฑ)๋กœ ์ด๋ฃจ์–ด์ ธ ์žˆ๋‹ค. Figure 1: ์ „์ฒด DRLN ๋„คํŠธ์›Œํฌ ๊ตฌ์กฐ.

์ „์ฒด ์†์‹ค ํ•จ์ˆ˜๋Š” N๊ฐœ์˜ batch์—์„œ ์ถœ๋ ฅ ์ด๋ฏธ์ง€ ํ–‰๋ ฌ๊ณผ ๋ผ๋ฒจ ์ด๋ฏธ์ง€ ํ–‰๋ ฌ์˜ ์ฐจ์ด๋ฅผ L1 norm์„ ํ†ตํ•ด ๊ณ„์‚ฐํ•˜๋Š”๋ฐ, ์ด๋Š” L1-์†์‹ค ํ•จ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋‹ค๋ฅธ SOTA ๋ฐฉ๋ฒ•๋“ค๊ณผ ์•Œ๊ณ ๋ฆฌ์ฆ˜ ์„ฑ๋Šฅ ๋น„๊ต๋ฅผ ์šฉ์ดํ•˜๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•จ์ด๋‹ค.

Figure 2: L1 ์†์‹คํ•จ์ˆ˜
Figure 2: Residual ๋ธ”๋ก ๊ตฌ์กฐ
Figure 3:

Residual ๋ธ”๋ก ์—ฐ์‡„ ์ง„ํ–‰ ๋ถ€๋ถ„์€ ์—ฌ๋Ÿฌ๊ฐœ์˜ ์—ฐ์‡„ ๋ธ”๋ก์œผ๋กœ ์ด๋ฃจ์–ด์ ธ ์žˆ๋‹ค. ๊ฐ ์—ฐ์‡„๋ธ”๋ก์€ ์ด์ „ ์ •๋ณด๋ฅผ ๊ฐ€์ ธ์˜ค๋Š” ์—ฌ๋Ÿฌ skip-connection๋“ค๊ณผ feature concatenation, ๊ทธ๋ฆฌ๊ณ  dense residual Laplacian module(DRLM)๋กœ ์ด๋ฃจ์–ด์ ธ ์žˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ๊ฐ DRLM ๋ธ”๋ก๋“ค์€ residual ์œ ๋‹›๋“ค์„ ์ด˜์ด˜ํžˆ ์—ฐ๊ฒฐํ•˜๋Š” ๋ถ€๋ถ„๊ณผ ์••์ถ•ํ•˜๋Š” ๋ถ€๋ถ„, ๊ทธ๋ฆฌ๊ณ  Laplacian pyramid attention ์œ ๋‹›์œผ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ๋‹ค.

Laplacian attention

Figure 4: Laplacian attention ๊ตฌ์กฐ

Laplacian attention์„ ์‚ฌ์šฉํ•˜๋ฉด ์ด๋ฏธ์ง€ ์ดˆํ•ด์ƒํ™”์— ํ•„์ˆ˜์ ์ธ feature๋“ค ์‚ฌ์ด์˜ ๊ด€๊ณ„๋ฅผ ๋” ์ž˜ ์ •์˜ํ•˜์—ฌ ํ•™์Šต ํšจ๊ณผ๋ฅผ ์ฆ๊ฐ€์‹œํ‚จ๋‹ค.

์ „์ฒด ์ด๋ฏธ์ง€์—์„œ์˜ ํ†ต๊ณ„๋ฅผ

์ž…๋ ฅ๋œ ์ €ํ•ด์ƒ๋„ ์ด๋ฏธ์ง€ ํ–‰๋ ฌ์„ x, ํ•™์Šตํ• ๋•Œ์˜ ์ž…๋ ฅ ์ด๋ฏธ์ง€์— ํ•ด๋‹นํ•˜๋Š” ๊ณ ํ•ด์ƒ๋„ ๋ผ๋ฒจ ์ด๋ฏธ์ง€ ํ–‰๋ ฌ์„ y, ์ถœ๋ ฅ๋˜๋Š” ์ดˆํ•ด์ƒํ™”๋œ ์ด๋ฏธ์ง€ ํ–‰๋ ฌ์„ \y^hat, convolution ๊ณ„์ธต์„ f, ๋น„์„ ํ˜• ํ™œ์„ฑํ™”ํ•จ์ˆ˜(ReLU)๋ฅผ ฯ„๋ผ๊ณ  ํ–ˆ์„๋•Œ, feature ์ถ”์ถœ์—์„œ์˜ convolution layer๋Š” f_0=H_f(x)

์ด ๋…ผ๋ฌธ์—์„œ๋Š” ๊ฐ ์—ฐ์‡„ ๋ธ”๋ก์— 3๊ฐœ์˜ DRLM์„ ์‚ฌ์šฉํ•˜์˜€๊ณ , ๊ฐ DRLM ์•ˆ์— 3๊ฐœ์˜ Residual ๋ธ”๋ก์„ 3x3

4. Experiment & Result

Experimental setup

ํ•™์Šต ๋ฐ์ดํ„ฐ์…‹: DIV2K, Flicker2K + ๋ฐ์ดํ„ฐ augmentation(๋žœ๋ค ํšŒ์ „) ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ์…‹: SET5, SET14, URBAN100, B100, MANGA109 + ์ด๋ฏธ์ง€ degradation(Bicubic, blur) ํ‰๊ฐ€ ๋ฐฉ๋ฒ•: YCbCr ์ƒ‰๊ณต๊ฐ„์—์„œ ๋ฐ๊ธฐ ์ฑ„๋„์— PSNR๊ณผ, SSIM ๋น„๊ต ๋น„๊ต๊ตฐ: SOTA CNN ์•Œ๊ณ ๋ฆฌ์ฆ˜ - SRCNN, FSRCNN, VDSR, SCN, SPMSR, LapSRN, MSLapSRN, MemNet, EDSR, SRMDNF, D-DBPN, IRCNN, RDN, RCAN, CARN

batch: 16 ์ €ํ•ด์ƒ๋„ ์ž…๋ ฅ ์ด๋ฏธ์ง€ ํฌ๊ธฐ: 48 X 48 ์ตœ์ ํ™”ํ•จ์ˆ˜: ADAM(ฮฒ1=0.9, ฮฒ2=0.999, ฮต=10e-08) Learning rate: ์ฒ˜์Œ์— 10e-04, ๋งค 2x10e05 ๋ฐ˜๋ณต๋งˆ๋‹ค ์ ˆ๋ฐ˜์œผ๋กœ ๊ฐ์†Œ ํ”„๋ ˆ์ž„์›Œํฌ: PyTorch GPU: Tesla P100

Result

Figure 5: Parameter ์ˆ˜ ๋Œ€๋น„ ์„ฑ๋Šฅ.
Figure 6:

5. Conclusion

๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ๋†’์€ ์ •ํ™•๋„์˜ ์ด๋ฏธ์ง€ ์ดˆํ•ด์ƒํ™” ์ž‘์—…์„ ์œ„ํ•œ ๋ชจ๋“ˆํ™”๋œ CNN์„ ์ œ์•ˆํ•˜์˜€๋‹ค. ์ดˆํ•ด์ƒํ™”์˜ ์„ฑ๋Šฅ์„ ๊ฐ•ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์—ฌ๋Ÿฌ ์š”์†Œ๋ฅผ ์ ์šฉํ•˜์˜€๋‹ค.

  1. Long skip connection, short skip connection, local connection์„ ์ด์šฉํ•ด residual ๊ตฌ์กฐ๋ฅผ ๋งŒ๋“ค๊ณ  ๊ทธ ๊ตฌ์กฐ๋ฅผ ์—ฐ์‡„์ ์œผ๋กœ ์ง„ํ–‰ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ฑ„์šฉํ•จ์œผ๋กœ์จ, ์ €ํ•ด์ƒ๋„์˜ ์ •๋ณด์˜ ํ๋ฆ„์„ ์ด์šฉํ•ด ๋„คํŠธ์›Œํฌ๊ฐ€ ๊ณ ํ•ด์ƒ๋„, ์ค‘๊ฐ„ ํ•ด์ƒ๋„์˜ ์ •๋ณด๋ฅผ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜์˜€๋‹ค.

  2. Feature๋ฅผ ๊ณ„์‚ฐํ•˜๋Š”๋ฐ ์‚ฌ์šฉํ•œ residual ๋ธ”๋ก์„ ๋ฐ€์ ‘ํ•˜๊ฒŒ ์—ฐ๊ฒฐํ•˜์—ฌ ์•”๋ฌต์ ์ธ "deep supervision"๊ณผ ๋†’์€ ๋‹จ๊ณ„์˜ ๋ณต์žกํ•œ feature๋“ค๋กœ๋ถ€ํ„ฐ์˜ ํ•™์Šต ๋“ฑ์˜ ์žฅ์ ์„ ๊ฐ–๊ฒŒ ๋˜์—ˆ๋‹ค.

  3. Laplacian attention์„ ์ด์šฉํ•˜์—ฌ ์—ฌ๋Ÿฌ ์Šค์ผ€์ผ์—์„œ์˜ ์ฃผ์š” feature๋ฅผ ์ƒ์„ฑํ•˜๊ณ  ๊ฐ feature map ์‚ฌ์ด์˜ ์˜์กด๋„๋ฅผ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•˜์˜€๋‹ค.

์„ค๊ณ„๋œ ๋„คํŠธ์›Œํฌ์— ๋Œ€ํ•ด ์ข…ํ•ฉ์ ์ธ ํ‰๊ฐ€๋ฅผ ๋ถ„์„ํ•˜์—ฌ ์„ฑ๋Šฅ์„ ์ž…์ฆํ•˜์˜€๋‹ค. ๋…ธ์ด์ฆˆ๋ฅผ ๊ฐ–๋Š” ์ €ํ•ด์ƒ๋„ ์ด๋ฏธ์ง€๋“ค๊ณผ unknown blur downsampling์„ ๊ฑฐ์นœ ์‹ค์ œ ์ด๋ฏธ์ง€๋ฅผ ํฌํ•จํ•œ ์ดˆํ•ด์ƒํ™” ๋ฐ์ดํ„ฐ์…‹์„ ์ด์šฉํ•˜์—ฌ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜์˜€๋‹ค. Bicubic ๋ฐ์ดํ„ฐ์…‹๊ณผ blur-down kernel์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ ํšจ์œจ์„ฑ์„ ์ž…์ฆํ•  ์ˆ˜ ์žˆ์—ˆ๊ณ , ๊ฐ๊ธฐ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•์œผ๋กœ ์ดˆํ•ด์ƒํ™”ํ•œ ์ด๋ฏธ์ง€๋“ค์— ๋Œ€ํ•ด ๊ฐ์ฒด์ธ์‹ ์„ฑ๋Šฅ์„ ๋ถ„์„ํ•˜์˜€๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ์ดˆํ•ด์ƒํ™”์— ๋Œ€ํ•œ DRLN ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ๋ถ„์„ํ–ˆ์ง€๋งŒ, ์‚ฌ์šฉ๋œ ๋ฐฉ๋ฒ•์ด ์ผ๋ฐ˜์ ์ด๊ธฐ์— ์ด๋ฏธ์ง€ ๋ณต์›, ํ•ฉ์„ฑ, ๋ณ€ํ™˜ ๋“ฑ์˜ ๋‹ค๋ฅธ low-level ๋น„์ „ ์ž‘์—…์—๋„ ์ ์šฉํ•  ์ˆ˜ ์žˆ์„ ๊ฒƒ์ด๋ผ ๊ธฐ๋Œ€ํ•œ๋‹ค. ์ ์šฉ๋œ ์†์‹ค ํ•จ์ˆ˜๋Š” L1 norm์„ ์ด์šฉํ•œ ๋ฐฉ๋ฒ•์ด์—ˆ๋Š”๋ฐ, ์ด๋Š” Residual ๋ธ”๋ก ์—ฐ์‡„ ์ง„ํ–‰๊ณผ Laplacian attention ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ์„ฑ๋Šฅ ๋ณ€ํ™”๋ฅผ ๋ณด๊ธฐ ์œ„ํ•จ์ด์—ˆ๋‹ค. ๋‹ค๋ฅธ ์†์‹ค ํ•จ์ˆ˜๋ฅผ ์ ์šฉํ–ˆ์„๋•Œ์˜ ๊ฒฐ๊ณผ๊ฐ€ ๊ธฐ๋Œ€๊ฐ€ ๋œ๋‹ค. ๋˜ํ•œ ์ž…๋ ฅ ์ด๋ฏธ์ง€๋ฅผ ๋‚˜๋ˆ„์–ด์„œ ๋ชจ๋ธ์„ ํ†ต๊ณผ์‹œํ‚ค๊ณ  ๋‹ค์‹œ ํ•ฉ์น˜๋Š” ์ž‘์—…์„ ํ•˜๋ฉด ์‹คํ–‰์†๋„์™€ ์ •ํ™•๋„์˜ ํ–ฅ์ƒ์ด ๋” ํด ๊ฒƒ์œผ๋กœ ๋ณด์ธ๋‹ค.

Take home message (์˜ค๋Š˜์˜ ๊ตํ›ˆ)

All men are mortal.

Socrates is a man.

Therefore, Socrates is mortal.

Author / Reviewer information

Author

์˜ค์ƒ์œค (Sangyoon Oh)

  • Affiliation (KAIST Mechanical Engineering)

  • E-mail: bulsajyo@kaist.ac.kr

Reviewer

Reference & Additional materials

  1. S. Anwar and N. Barnes, "Densely Residual Laplacian Super-Resolution" in IEEE Transactions on Pattern Analysis & Machine Intelligence, vol. , no. 01, pp. 1-1, 5555.

  2. https://github.com/saeed-anwar/DRLN

  3. Citation of related work

  4. Other useful materials

  5. ...

Last updated

Was this helpful?