Mobile QR Code QR CODE : The Transactions P of the Korean Institute of Electrical Engineers

  1. (Dept. of AI Convergence Engineering, Gyeongsang National University, Korea.)



Machine Learning, Generative Adversarial Network, Line Arts Colorization, Image Generation

1. ์„œ ๋ก 

DCGAN(Deep Convolutional Generative Adversarial Network) [1]์ด ๋ฐœํ‘œ๋จ์— ๋”ฐ๋ผ GAN์˜ ์—ฐ๊ตฌ๊ฐ€ ํ™œ๋ฐœํžˆ ์ง„ํ–‰๋˜๊ณ  ์žˆ๋‹ค. GAN(Generative Adversarial Network) ์—ฐ๊ตฌ๊ฐ€ ์ง„ํ–‰๋จ์— ๋”ฐ๋ผ ์—ฌ๋Ÿฌ ์‘์šฉ ๋ถ„์•ผ, ํŠนํžˆ ์ด๋ฏธ์ง€์ƒ์„ฑ์—์„œ ๋›ฐ์–ด๋‚œ ๊ฒฐ๊ณผ๊ฐ€ ํ™œ๋ฐœํžˆ ๋‚˜์˜ค๊ณ  ์žˆ๋‹ค. ์„ ํ™”๋Š” ๋‹ค์–‘ํ•œ ๋ฏธ๋””์–ด ์‚ฐ์—…์— ์ž‘ํ’ˆ ๋ฐฉํ–ฅ์„ ๊ฒฐ์ •ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ ๋˜๋ฉฐ, ํ”„๋กœ์ ํŠธ ์ดˆ๊ธฐ ๋ฐฉํ–ฅ์„ ๊ฒฐ์ •ํ•˜๋Š” ๋ฐ ๋งค์šฐ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•œ๋‹ค. ์„ ํ™”์˜ ์ฑ„์ƒ‰์€ ์ˆ™๋ จ๋œ ์•„ํ‹ฐ์ŠคํŠธ๊ฐ€ ์ „๋ฌธ ํŽธ์ง‘ ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•ด ์ง„ํ–‰ํ•œ๋‹ค. ๋‹ค์–‘ํ•œ ๋ ˆ์ด์–ด๋ฅผ ์‚ฌ์šฉํ•ด ์ฑ„์ƒ‰ํ•˜๋Š” ์ž‘์—…์€ ๋…ธ๋™์ง‘์•ฝ์ ์ด๊ณ  ์ง€๋ฃจํ•œ ๋ฐ˜๋ณต ์ž‘์—…์„ ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— GAN์„ ์‚ฌ์šฉํ•ด ์„ ํ™”๋ฅผ ์ฑ„์ƒ‰ํ•˜๋Š” ์—ฐ๊ตฌ๊ฐ€ ์ง„ํ–‰๋˜๊ณ  ์žˆ์œผ๋ฉฐ Petalica Paint (๊ตฌ Paints Chainer) [2] ์™€ ๊ฐ™์€ ์ƒ์šฉํ™”๋œ ์ž๋™์ฑ„์ƒ‰ ๋„๊ตฌ๋“ค์ด ์‚ฌ์šฉ๋˜๊ณ  ์žˆ๋‹ค.

๊ธฐ์กด ์„ ํ™” ์ž๋™์ฑ„์ƒ‰์€ ์ž…๋ ฅํ•˜๋Š” ๋ฐ์ดํ„ฐ์— ๋”ฐ๋ผ ํฌ๊ฒŒ ์„ธ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์ด ์žˆ๋‹ค. ์ฒซ์งธ, ์„ ํ™”๋งŒ ์‚ฌ์šฉํ•ด ์ฑ„์ƒ‰๋œ ์ด๋ฏธ์ง€๋ฅผ ์ƒ์„ฑํ•˜๋Š” ์™„์ „ ์ž๋™๋ฐฉ์‹ [3,4], ๋‘˜์งธ, ์„ ํ™”์™€ ์ปฌ๋Ÿฌ ์ด๋ฏธ์ง€๋ฅผ ์‚ฌ์šฉํ•ด ์„ ํ™”๋ฅผ ์ž…๋ ฅ ํ•œ ์ปฌ๋Ÿฌ ์ด๋ฏธ์ง€ ์Šคํƒ€์ผ๋กœ ์ฑ„์ƒ‰ํ•˜๋Š” ์Šคํƒ€์ผ ๋ณ€ํ™˜์„ ํ†ตํ•œ ์ž๋™๋ฐฉ์‹ [5,6,7], ์…‹์งธ, ์„ ํ™”์™€ ์‚ฌ์šฉ์ž ํžŒํŠธ๋ฅผ ์ž…๋ ฅํ•ด ์›ํ•˜๋Š” ์ƒ‰์œผ๋กœ ์ฑ„์ƒ‰ํ•˜๋Š” ๋ฐฉ์‹์ด๋‹ค. [8, 9, 10, 11, 12, 13, 14]

์ž๋™์ฑ„์ƒ‰ ๋ถ„์•ผ์—์„œ ์ฑ„์ƒ‰ ๋ชจ๋ธ์— ๊ด€ํ•œ ์—ฐ๊ตฌ๋Š” ๋‹ค์–‘ํ•˜๊ฒŒ ๋ฐœํ‘œ๋˜๊ณ  ์žˆ์ง€๋งŒ, ์ฑ„์ƒ‰ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜๊ธฐ ์œ„ํ•œ ์„œ๋น„์Šค ๊ด€์ ์˜ ์—ฐ๊ตฌ๋Š” ๋ณด์ด์ง€ ์•Š๊ณ  ์žˆ๋‹ค. ๋˜ํ•œ ์ƒ์šฉํ™”๋˜์ง€ ์•Š์€ ์„œ๋น„์Šค์˜ ๊ฒฝ์šฐ ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•œ ํ…Œ์ŠคํŠธ๋กœ ํ˜„์žฌ๋Š” ํ™œ์„ฑํ™”๋˜์–ด ์žˆ์ง€ ์•Š๋‹ค. ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” Lee et al. [14]๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์‚ฌ์šฉ์ž๋“ค์ด ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์•ž์œผ๋กœ ์ž๋™์ฑ„์ƒ‰์—ฐ๊ตฌ์— ํ…Œ์ŠคํŠธ๋ฅผ ์œ„ํ•œ ํ”Œ๋žซํผ์œผ๋กœ์จ ํ™œ์šฉ ๊ฐ€๋Šฅํ•œ ์„ ํ™” ์ž๋™ ์ฑ„์ƒ‰ ์›น์„œ๋น„์Šค๋ฅผ ์ œ์•ˆํ•œ๋‹ค.

๋ณธ ์—ฐ๊ตฌ๊ฐ€ ์ œ์•ˆํ•˜๋Š” ์„œ๋น„์Šค๋Š” ONNX (Open Neural Network Exchange) [15]๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ, CPU ํ™˜๊ฒฝ์—์„œ ์ถ”๋ก ์„ ์ง€์›ํ•˜๊ณ  ๋‹ค์–‘ํ•œ ์ž…์ถœ๋ ฅ ๋ชจ๋ธ์„ ์œ„ํ•œ ๊ณ ์ฐจํ•จ์ˆ˜ ๊ธฐ๋ฐ˜์˜ ์ „์ฒ˜๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค. ์„œ๋น„์Šค๋Š” ์ฑ„์ƒ‰๋ชจ๋ธ์„ ํ•™์Šต ๋ฐ ๋ณ€ํ™˜ํ•˜๋Š” ๋ชจ๋ธ ์ƒ์„ฑ๊ธฐ, ์ „์ฒ˜๋ฆฌ์™€ ์ƒ์„ฑ ๋ชจ๋ธ์˜ ๊ด€๋ฆฌ ๋ฐ ์ถ”๋ก ์„ ์ง„ํ–‰ํ•˜๋Š” ์ถ”๋ก  ์„œ๋ฒ„, ์‚ฌ์šฉ์ž๋ฅผ ์œ„ํ•œ ์„œ๋น„์Šค ํ”„๋ŸฐํŠธ ์—”๋“œ๋กœ ๊ตฌ์„ฑ๋œ๋‹ค. ์ œ์ž‘๋œ ์„œ๋น„์Šค์˜ ํ˜•์ƒ์€ ๊ทธ๋ฆผ 1๊ณผ ๊ฐ™๋‹ค. ์ œ์•ˆํ•œ ์„œ๋น„์Šค๋Š” CPU ํ™˜๊ฒฝ์—์„œ Pytorch์˜ 2.2683 ์ดˆ์™€ ๋น„๊ตํ•ด ์žฅ๋‹น ํ‰๊ท  0.4040 ์ดˆ๋กœ 5๋ฐฐ ๋น ๋ฅธ ์†๋„๋กœ ํšจ์œจ์ ์ธ ์ถ”๋ก ์ด ๊ฐ€๋Šฅํ–ˆ๋‹ค.

๋ณธ ๋…ผ๋ฌธ์˜ ๊ตฌ์„ฑ์€ ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค. 2์žฅ์—์„œ๋Š” ๊ธฐ์กด ์ƒ์šฉํ™”๋œ ์ž๋™ ์ฑ„์ƒ‰ ์„œ๋น„์Šค์™€ ์ž…๋ ฅ์— ๋”ฐ๋ผ ๊ธฐ์กด ์ž๋™์ฑ„์ƒ‰ ์—ฐ๊ตฌ๋ฅผ ๋ถ„๋ฅ˜ ๋ฐ ์„ค๋ช…ํ•œ๋‹ค. 3์žฅ์—์„œ๋Š” ์ œ์•ˆํ•˜๋Š” ์„œ๋น„์Šค์— ์‚ฌ์šฉ๋œ ์ž๋™์ฑ„์ƒ‰ ๋ชจ๋ธ์„ ํ•™์Šตํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ์šฉ๋œ ํ•ต์‹ฌ ๊ธฐ๋ฒ•์„ ์„ค๋ช…ํ•œ๋‹ค. 4์žฅ์—์„œ๋Š” ์ž๋™์ฑ„์ƒ‰ ์„œ๋น„์Šค ํ”Œ๋žซํผ์˜ ๊ตฌ์„ฑ์š”์†Œ์— ๋Œ€ํ•ด ์„ค๋ช…ํ•œ๋‹ค. 5์žฅ์—์„œ๋Š” ์ถ”๋ก ๋‹จ๊ณ„์—์„œ ์‚ฌ์šฉ๋œ ONNX์˜ ์„ฑ๋Šฅ๋น„๊ต๋ฅผ ์œ„ํ•œ ์‹คํ—˜ ๋ฐ ๊ธฐ์กด ์„œ๋น„์Šค์™€์˜ ๊ฒฐ๊ณผ๋ฌผ์„ ์‹œ๊ฐ์ ์œผ๋กœ ๋น„๊ต ๋ฐ ๋ถ„์„ํ•œ ๊ฒฐ๊ณผ๋ฅผ ์„ค๋ช…ํ•œ๋‹ค. 6์žฅ์—์„œ๋Š” ๋…ผ๋ฌธ์˜ ๊ฒฐ๋ก ์„ ์ œ์‹œํ•œ๋‹ค.

๊ทธ๋ฆผ 1. ์„œ๋น„์Šค ํ˜•์ƒ

Fig. 1. Service View

../../Resources/kiee/KIEEP.2022.71.1.41/fig1.png

2. ๊ด€๋ จ ์—ฐ๊ตฌ

์ด ์žฅ์—์„œ๋Š” ์ƒ์šฉํ™”๋˜์–ด ์žˆ๋Š” ์ž๋™์ฑ„์ƒ‰ ์„œ๋น„์Šค ๊ทธ๋ฆฌ๊ณ  ์‹ ๊ฒฝ๋ง์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋Š” ์„ ํ™” ์ž๋™์ฑ„์ƒ‰ ๊ธฐ๋ฒ•์„ ์ž…๋ ฅ ๋ฐ์ดํ„ฐ์— ๋”ฐ๋ผ ๋ถ„๋ฅ˜ํ•˜๊ณ  ์„ค๋ช…ํ•œ๋‹ค. ์„ ํ™” ์ž๋™์ฑ„์ƒ‰์€ ์ž…๋ ฅ ๋ฐ์ดํ„ฐ ํ˜•์‹์— ๋”ฐ๋ผ ์„ธ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์œผ๋กœ ๋‚˜๋ˆ„์–ด์ง„๋‹ค. ๊ฐ ๋ฐฉ๋ฒ•์œผ๋กœ๋Š” ์™„์ „ ์ž๋™ ์ฑ„์ƒ‰ ๋ฐฉ์‹, ์Šคํƒ€์ผ ๋ณ€ํ™˜์„ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ์‹, ์‚ฌ์šฉ์ž ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”๊ฐ€๋กœ ์ž…๋ ฅํ•˜๋Š” ๋ฐฉ์‹์ด ์žˆ๋‹ค. [3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14]

2.1 Line-Arts Automatic Colorization Service

๊ธฐ์กด ์„ ํ™” ์ž๋™์ฑ„์ƒ‰ ์„œ๋น„์Šค๋กœ๋Š” Petalica Paint [2] ๋ฐ Clip Studio๊ฐ€ ์žˆ๋‹ค. Petalica Paint๋Š” ์˜จ๋ผ์ธ ์„œ๋น„์Šค๋ฅผ ํ†ตํ•ด 512 x 512 pixel ํ•ด์ƒ๋„ ๊ทœ๊ฒฉ์œผ๋กœ ์ฑ„์ƒ‰์„ ์ง€์›ํ•˜๋ฉฐ ์ฑ„์ƒ‰ ํŠน์ง•์— ๋”ฐ๋ผ โ€œTanpopoโ€, โ€œSatsukiโ€, โ€œCannaโ€ 3๊ฐ€์ง€์˜ ๋ชจ๋ธ์„ ์ง€์›ํ•œ๋‹ค. Clip Studio๋Š” Photoshop ๊ณผ ๊ฐ™์€ ๋ ˆ์ด์–ด๋ฅผ ์ง€์›ํ•˜๋Š” ์ƒ์—…์šฉ ๋„๊ตฌ์ด๋ฉฐ ์„  ๋ ˆ์ด์–ด์™€ ๋ณ„๋„์˜ ํžŒํŠธ ๋ ˆ์ด์–ด๋ฅผ ๋งŒ๋“ค์–ด ์ฑ„์ƒ‰ ์ด๋ฏธ์ง€๊ฐ€ ์ƒ์„ฑ๋œ ๋ ˆ์ด์–ด๋ฅผ ๋งŒ๋“ค์–ด ๋‚ธ๋‹ค. Clip Studio๋Š” ๋ ˆ์ด์–ด๋ฅผ ์ง€์›ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์„  ์˜์—ญ ์™ธ๋ถ€๋กœ ๋ฒˆ์ง„ ์ƒ‰์ƒ ์ด๋ฏธ์ง€๋ฅผ ๋งˆ์Šคํ‚นํ•˜๋Š” ๋ฐฉ์‹์„ ํ†ตํ•ด ๊ณ ํ’ˆ์งˆ์˜ ๊ฒฐ๊ณผ๋ฌผ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ์žฅ์ ์ด ์žˆ๋‹ค.

2.2 Fully Automatic Colorization

์™„์ „์ž๋™๋ฐฉ์‹์˜ ์ฑ„์ƒ‰ ๊ธฐ๋ฒ•[3,4]์€ ๋‹ค๋ฅธ ํ˜•ํƒœ์˜ ์ž…๋ ฅ ์—†์ด ์„ ํ™”๋งŒ์„ ์‚ฌ์šฉํ•œ๋‹ค. Isola et al. [3] (Pix2Pix)๋Š” ์—ฐ๊ตฌ [9] ์˜ ์กฐ๊ฑด์ž…๋ ฅ์„ ์‚ฌ์šฉํ•œ cGAN(Conditional Generative Adversarial Networks) ๊ตฌ์กฐ ๋กœ ์ด๋ฏธ์ง€ ๋Œ€ ์ด๋ฏธ์ง€ ๋ณ€ํ™˜์˜ ์†”๋ฃจ์…˜์„ ์ œ๊ณตํ–ˆ๋‹ค. ์—ฐ๊ตฌ [3]์€ ์‚ฌ์‹ค ์ ์ธ ์ด๋ฏธ์ง€๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด L1 ์†์‹ค ๋ฐ ์ ๋Œ€์  ์†์‹ค์„ ๊ฒฐํ•ฉํ•ด L1 ์†์‹ค๋งŒ ์‚ฌ์šฉํ•œ ๊ฒฐ๊ณผ์— ๋น„ํ•ด ์„ ๋ช…ํ•˜๊ณ  ์‚ฌ์‹ค์ ์ธ(photorealistic) ์ด๋ฏธ์ง€๋ฅผ ์ƒ์„ฑํ–ˆ๋‹ค. Kang et al. [4] ์—์„œ๋Š” ์ฑ„์ƒ‰์ž‘์—…์„ ์œ„ํ•œ 3๊ฐ€์ง€ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•œ๋‹ค. ๊ฐ๊ฐ์˜ ๋ชจ๋ธ์€ ์‹ค์งˆ์ ์ธ ์ฑ„์ƒ‰์„ ๋‹ด๋‹นํ•˜๋Š” โ€œLow-resolution Colorizerโ€, ์ „๊ฒฝ๊ณผ ๋ฐฐ๊ฒฝ์„ ๋ถ„๋ฅ˜ํ•˜๋Š” โ€œBackground Detectorโ€, ๊ทธ๋ฆฌ๊ณ  ์ฑ„์ƒ‰๋œ ์ €ํ•ด์ƒ๋„ ์ด๋ฏธ์ง€์™€ ๋ฐฐ๊ฒฝ Segment๋ฅผ ๋ฐ›์•„ ๋ฐฐ๊ฒฝ์„ ๊ตฌ๋ถ„ํ•˜์—ฌ ํ•ด์ƒ๋„๋ฅผ ๋ณต์›ํ•˜๋Š” โ€œPolishing Networkโ€๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค. ์ด ์—ฐ๊ตฌ [4] ์€ ๋งํ’์„ ๊ณผ ๊ฐ™์€ ๋งŒํ™”์˜ ํŠน์ง•์„ ์ž˜ ํ™œ์šฉํ•˜์˜€๊ณ  ์„ ํ™”๋ฅผ ์ผ๊ด„์ ์œผ๋กœ ์ฑ„์ƒ‰ํ•  ์ˆ˜ ์žˆ๋Š” ์žฅ์ ์ด ์žˆ๋‹ค. ํ•˜์ง€๋งŒ, ์™„์ „ ์ž๋™์œผ๋กœ ์ฑ„์ƒ‰์„ ์ง„ํ–‰ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์›ํ•˜๋Š” ๋ถ€์œ„๋ฅผ ํŠน์ •ํ•œ ์ƒ‰์œผ๋กœ ์ฑ„์ƒ‰ํ•˜๊ธฐ๋Š” ๋ถˆ๊ฐ€๋Šฅํ•˜๊ณ  ์ถœ๋ ฅ ์ด๋ฏธ์ง€์˜ ํฌ๊ธฐ๊ฐ€ 256 x 256 pixel ํ•ด์ƒ๋„๋กœ ํ•œ์ •๋˜๋Š” ๋‹จ์ ์ด ์žˆ๋‹ค.

2.3 Style Transfer based Colorization

์ผ๋Ÿฌ์ŠคํŠธ ์ž๋™์ฑ„์ƒ‰์„ ์œ„ํ•œ ๋˜ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ์Šคํƒ€์ผ ๋ณ€ํ™˜ ๊ธฐ๋ฐ˜์˜ ์—ฐ๊ตฌ๊ฐ€ ์žˆ๋‹ค [5,6,7]. ์Šคํƒ€์ผ ๋ณ€ํ™˜์˜ ๊ฒฝ์šฐ ์„ ํ™” ์ด๋ฏธ์ง€์™€ ์ฐธ๊ณ ํ•  ์Šคํƒ€์ผ์ด ๋˜๋Š” ์ปฌ๋Ÿฌ ์ด๋ฏธ์ง€ 2๊ฐœ์˜ ์ž…๋ ฅ์„ ๊ธฐ๋ณธ์œผ๋กœ ์‚ฌ์šฉํ•œ๋‹ค. Furusawa et al. [5]์€ ์ฐธ์กฐ ์ด๋ฏธ์ง€์™€ ๋Œ€ํ™”์‹ ์ƒ‰์ƒ ํžŒํŠธ(์ƒ‰ ํŒ”๋ ˆํŠธ) ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์‚ฌ์šฉํ•ด ๋ฐ˜์ž๋™์œผ๋กœ ์‚ฌ์šฉ์ž๊ฐ€ ์›ํ•˜๋Š” ์ƒ‰์„ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜์˜€์œผ๋ฉฐ ์„ ํ™”๊ฐ€ ์•„๋‹Œ ๋งŒํ™” ์ฝค๋งˆ ์ด๋ฏธ์ง€๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ฑ„์ƒ‰์„ ์ง„ํ–‰ํ•˜์˜€๋‹ค. ์ฑ„์ƒ‰ ๊ณผ์ •์—์„œ ์ƒ‰ ์ •๋ณด๋ฅผ ์ƒ์„ฑํ•˜๊ณ  ์›๋ณธ ๋งŒํ™” ์ด๋ฏธ์ง€์—์„œ ์œค๊ณฝ์„ ์„ ์ถ”์ถœํ•˜์—ฌ ํ•ฉ์„ฑํ•˜๋Š” ๊ตฌ์กฐ๋กœ ํšจ์œจ์ ์œผ๋กœ ์ฑ„์ƒ‰ํ•˜์˜€๋‹ค. ํ•˜์ง€๋งŒ, ์‚ฌ์šฉ์ž์˜ ์ƒ‰์ƒ ์ •๋ณด๊ฐ€ ์ง๊ด€์ ์œผ๋กœ ์›ํ•˜๋Š” ์œ„์น˜์— ๋“ค์–ด๊ฐ€์ง€ ์•Š์œผ๋ฉฐ ์ƒ‰ ์ •๋ณด์™€ ํ…์ŠคํŠธ์™€ ๊ฐ™์€ ์œค๊ณฝ์„ ์„ ์œ„์—์„œ ๋ฎ์–ด์“ฐ๋Š” ๊ตฌ์กฐ๋กœ ์ธํ•ด ์ด๋ฏธ์ง€ ์งˆ๊ฐ์— ์†์ƒ์ด ์‹ฌํ•˜๋‹ค๋Š” ๋‹จ์ ์ด ์žˆ๋‹ค. Zhang et al. [7] ์€ VGG16/19 ๊ตฌ์กฐ [16]์˜ ๋ชจ๋ธ์„ ํ†ตํ•ด ์Šคํƒ€์ผ ์ด๋ฏธ์ง€๋ฅผ ์ถ”๊ฐ€ํ•ด ์ฑ„์ƒ‰์„ ์ง„ํ–‰ํ–ˆ๋‹ค. ๋ชจ๋ธ ์ค‘๊ฐ„์˜ ๋‘ ๊ฐœ์˜ โ€œGuide Decoderโ€ ์‚ฌ์šฉํ•จ์œผ๋กœ ํ•™์Šต์—์„œ์˜ ๊ธฐ์šธ๊ธฐ๊ฐ€ ์‚ฌ๋ผ์ง€๋Š” ๋ฌธ์ œ (Vanishing Gradient)๋ฅผ ๋ฐฉ์ง€ ํ•˜๊ณ  ํšจ๊ณผ์ ์ธ ์ฑ„์ƒ‰์„ ํ–ˆ๋‹ค. ํ•˜์ง€๋งŒ, VGG16/19 ๋„คํŠธ์›Œํฌ๋ฅผ ์‚ฌ์šฉํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋„คํŠธ์›Œํฌ๋ชจ๋ธ ์šฉ๋Ÿ‰์ด ํฌ๊ณ , ์ž๋™์œผ๋กœ ์ฑ„์ƒ‰ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์›ํ•˜๋Š” ๋ถ€์œ„์— ์›ํ•˜๋Š” ์ƒ‰์œผ๋กœ ์ฑ„์ƒ‰ํ•˜๊ธฐ ํž˜๋“ค๋ฉฐ, ์ถœ๋ ฅ ์ด๋ฏธ์ง€์˜ ํฌ๊ธฐ๊ฐ€ 256x256 pixel ํ•ด์ƒ๋„๋กœ ํ•œ์ •๋œ๋‹ค๋Š” ๋‹จ์ ์ด ์žˆ๋‹ค.

2.4 Colorization with Color Point Hinting

๋งˆ์ง€๋ง‰์œผ๋กœ ์„ ํ™” ์ด๋ฏธ์ง€์— ์‚ฌ์šฉ์ž ์ปฌ๋Ÿฌ ํžŒํŠธ๋ฅผ ์ถ”๊ฐ€๋กœ ์ž…๋ ฅํ•˜์—ฌ ์ปฌ๋Ÿฌ ํžŒํŠธ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ฑ„์ƒ‰ํ•˜๋Š” ์—ฐ๊ตฌ๊ฐ€ ์žˆ๋‹ค [8, 9, 10, 11, 12, 13]. ์ปฌ๋Ÿฌ ์„ ์„ ํžŒํŠธ๋กœ ์‚ฌ์šฉํ•˜๋Š” ์—ฐ๊ตฌ ์ค‘ ๋Œ€ํ‘œ์ ์œผ๋กœ Ci et al. [11] ์ด ์žˆ๋‹ค. ์ด ์—ฐ๊ตฌ [11]์—์„œ๋Š” ๋ชจ๋ธ์˜ ์ธ๊ณต์„ ํ™”(์›๋ณธ ์ปฌ๋Ÿฌ ์ด๋ฏธ์ง€์—์„œ ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ ๋งŒ๋“ค์–ด๋‚ธ ์„ ํ™”)์˜ ๊ณผ์ ํ•ฉ(overfitting)์„ ๋ง‰๊ธฐ ์œ„ํ•ด LFN(Local Feature Net)์„ ์‚ฌ์šฉํ•œ๋‹ค. LFN์„ ํ†ตํ•ด ์ž…๋ ฅ ์„ ํ™” ํŠน์ง•์„ ์ถ”์ถœํ•ด ์ƒ์„ฑ์ž์™€ ๋ถ„๋ฅ˜์ž์˜ ์ถ”๊ฐ€ ์ž…๋ ฅ์œผ๋กœ ์‚ฌ์šฉํ•ด ๊ณผ์ ํ•ฉ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ–ˆ๋‹ค๊ณ  ํ•œ๋‹ค. ํ•˜์ง€๋งŒ, Loss ๊ณ„์‚ฐ ์‹œ VGG16 ๋„คํŠธ์›Œํฌ๋ฅผ ์‚ฌ์šฉํ•˜๋ฏ€๋กœ ๋ชจ๋ธ ์šฉ๋Ÿ‰์ด ํฌ๋‹ค๋Š” ๋‹จ์ ์ด ์žˆ๋‹ค. Sangkloy et al. [8]์€ ์ธ๊ณต ์„ ํ™”์˜ ๊ณผ์ ํ•ฉ์„ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•ด 4๊ฐ€์ง€์˜ ๋‹ค๋ฅธ ํ•„ํ„ฐ๋ฅผ ์ ์šฉํ•ด ์„ ํ™”๋ฅผ ์ถ”์ถœํ•˜์—ฌ ์–ผ๊ตด ์ฑ„์ƒ‰์„ ์ง„ํ–‰ํ•˜์˜€๋‹ค. Frans et al. [10]์€ ์ž…๋ ฅ ์„ ํ™”์—์„œ ์ฑ„์ƒ‰ ๋ฐ ์ฑ„์ƒ‰ ๊ฒฐ๊ณผ๋ฌผ์˜ ์Œ์˜์„ ์ƒ์„ฑํ•˜๋Š” ์ƒ์„ฑ์ž๋ฅผ ๋ณ„๋„๋กœ ํ•™์Šตํ•˜์—ฌ 2์ค‘ ์ƒ์„ฑ์ž๋ฅผ ์‚ฌ์šฉํ•ด ํšจ๊ณผ์ ์ธ ์ฑ„์ƒ‰๊ธฐ๋ฒ•์„ ๋ณด์˜€๋‹ค. ์ปฌ๋Ÿฌ ์ ์„ ์‚ฌ์šฉํ•œ ์—ฐ๊ตฌ๋กœ๋Š” Liu et al. [9]์ด ์žˆ์œผ๋ฉฐ ์ƒ์„ฑ์ž ํ•™์Šต์„ ์œ„ํ•œ Loss๋ฅผ ๋‚˜๋ˆ„์–ด ๊ฐ๊ฐ์˜ Loss ๊ณ„์ˆ˜ ํ•ญ์„ ์กฐ์ ˆํ•˜์—ฌ ํ•™์Šต์„ ์ง„ํ–‰ํ•˜์˜€๋‹ค. ํ•™์Šต ๊ฒฐ๊ณผ, ์ƒ‰ ๋ฒˆ์ง์„ ํšจ๊ณผ์ ์œผ๋กœ ๋ฐฉ์ง€ํ•˜๋ฉด์„œ Pix2pix [3] ๋ชจ๋ธ ๋ณด๋‹ค ์ข‹์€ ์ด๋ฏธ์ง€๋ฅผ ๋ณด์—ฌ์ฃผ์—ˆ๋‹ค. ์ด์ค‘ ์ƒ์„ฑ์ž๋ฅผ ์ด์šฉํ•œ ๋‹ค๋ฅธ ์—ฐ๊ตฌ ์ธ HATI et al. [13]์€ Ci et al. [11]์˜ ๋ชจ๋ธ์˜ ๊ธฐ๋ฐ˜์œผ๋กœ ์ด์ค‘ ์ƒ์„ฑ์ž๋ฅผ ์‚ฌ์šฉํ•˜์˜€๋‹ค. ์ด์ค‘ ์ƒ์„ฑ์ž๋Š” ์ƒ์„ฑ๋œ ์ดˆ์•ˆ ์ด๋ฏธ์ง€๋กœ๋ถ€ํ„ฐ ํ•ฉ์„ฑ ์„  ํ™”๋ฅผ ์ถ”์ถœํ•œ๋‹ค. ๊ทธ๋ฆฌ๊ณ , ์ƒ์„ฑ๋œ ํ•ฉ์„ฑ ์„ ํ™”์™€ ์ดˆ์•ˆ ์ƒ์„ฑ ์‹œ ์‚ฌ์šฉํ•œ ์ธ๊ณต ์„ ํ™”์˜ ์žฌ๊ตฌ์„ฑ ์†์‹ค(reconstruction loss)์„ ๋ฐ˜์˜ํ•ด 1๋‹จ๊ณ„ ์ƒ์„ฑ์ž์˜ ์„ฑ๋Šฅ์„ ๋†’์˜€๋‹ค. Zhang et al. [12]์€ ์ด์ค‘ ์ƒ์„ฑ์ž ๊ตฌ์กฐ์—์„œ ํ›„์ฒ˜๋ฆฌ ๋ชจ๋ธ์˜ ์ดˆ์•ˆ ์˜์กด๋„๋ฅผ ์ค„์ด๊ธฐ ์œ„ํ•ด ์ƒ์„ฑ๋œ ์ดˆ์•ˆ ์ด๋ฏธ์ง€์— ์ƒ‰ ๋ฒˆ์ง ๋“ฑ๊ณผ ๊ฐ™์€ ์‹œ๋ฎฌ๋ ˆ์ด์…˜์„ ์ ์šฉํ•ด ํ›„ ์ฒ˜๋ฆฌ ๋ชจ๋ธ์˜ ์ดˆ์•ˆ ์˜์กด๋„๋ฅผ ์ค„์ด๋Š” ๋ฐฉ์‹์œผ๋กœ ์ฑ„์ƒ‰ ์„ฑ๋Šฅ์„ ๋†’์˜€๋‹ค.

3. ์ œ์•ˆ ๊ธฐ๋ฒ•์˜ ๊ตฌ์„ฑ

์ด ์žฅ์—์„œ๋Š” ์ œ์•ˆํ•˜๋Š” ์„œ๋น„์Šค์— ์‚ฌ์šฉ๋œ ๋ชจ๋ธ๊ณผ ํ•™์Šต์— ์‚ฌ์šฉ๋œ ํ•ต์‹ฌ ๊ธฐ๋ฒ•์— ๊ด€ํ•ด ์„ค๋ช…ํ•œ๋‹ค. ์ œ์•ˆํ•˜๋Š” ์„œ๋น„์Šค์˜ ๋ชจ๋ธ ๊ตฌ์กฐ์™€ ํ•™์Šต์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์„ค๋ช…์€ Lee et al. [14]์— ์ƒ์„ธํžˆ ์„ค๋ช…๋˜์–ด ์žˆ๋‹ค. ๋ณธ ์—ฐ๊ตฌ์—์„œ ์„œ๋น„์Šค๋ฅผ ๊ตฌํ˜„ํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ์šฉ๋œ ๊ธฐ๋ฒ•์„ ์„ค๋ช…ํ•œ๋‹ค. ๋ชจ๋ธ์— ์‚ฌ์šฉ๋œ ํ•ต์‹ฌ ํ•™์Šต ๊ธฐ๋ฒ•์€ ํžˆ์Šคํ† ๊ทธ๋žจ ํ‰ํ™œํ™”๋ฅผ ์‚ฌ์šฉํ•œ ์„ ํ™” ๋ฐ์ดํ„ฐ ์ฆ์‹(data augmentation), ์ดˆ์•ˆ๊ณผ ์ฑ„์ƒ‰ ์ž‘์—…์„ ๋ถ„๋ฆฌํ•œ ์ด์ค‘ ์ƒ์„ฑ์ž, ์„ ํ™” ํƒ์ง€ ๋ชจ๋ธ (Line Detection Model)์„ ์‚ฌ์šฉํ•œ ์„ ํ™” ์˜ค์ฐจํ•จ์ˆ˜๊ฐ€ ์žˆ๋‹ค.

3.1 ์„ ํ™” ๋ฐ์ดํ„ฐ ์ฆ์‹

์„ ํ™” ์ž๋™์ฑ„์ƒ‰์— ์žˆ์–ด ๊ฐ€์žฅ ํฐ ๋ฌธ์ œ๋Š” ๋‹ค์–‘ํ•œ ์„ ํ™” ๊ตฌ์กฐ์—์„œ ์•ˆ์ •์ ์ธ ์ฑ„์ƒ‰์ด ๋ถˆ๊ฐ€๋Šฅํ•œ ๊ฒƒ์ด๋‹ค. ์„ ํ™” ์ž๋™์ฑ„์ƒ‰ ๋ชจ๋ธ ํ•™์Šต์„ ์œ„ํ•ด์„œ๋Š” ์„ ํ™”์™€ ์ปฌ๋Ÿฌ ์Œ์œผ๋กœ ๊ตฌ์„ฑ๋œ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ ์…‹์ด ํ•„์š”ํ•˜๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ ์Œ์„ ๋Œ€๋Ÿ‰์œผ๋กœ ์ˆ˜์ง‘ํ•˜๋Š” ๊ฒƒ์€ ์–ด๋ ต๊ธฐ ๋•Œ๋ฌธ์— ๋Œ€๋ถ€๋ถ„ ์˜ ์—ฐ๊ตฌ์—์„œ๋Š” ์ปฌ๋Ÿฌ ์ด๋ฏธ์ง€์—์„œ ์„ ํ™” ์ด๋ฏธ์ง€๋ฅผ ์ถ”์ถœํ•˜์—ฌ ์ž๊ฐ€ ํ•™์Šต(Self-learning) ๋ฐฉ์‹์œผ๋กœ ํ•™์Šต์„ ์ง„ํ–‰ํ•œ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์€ ์ถ”์ถœ ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ๋”ฐ๋ผ ํŠน์ •ํ•œ ๋ถ„ํฌ๋ฅผ ๊ฐ€์ง€๋ฏ€๋กœ ํ•™์Šต ๊ณผ์ •์—์„œ ๊ณผ์ ํ•ฉ (overfitting)์ด ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ๋‹ค. ๋ณธ ์—ฐ๊ตฌ๋Š” Lee et al. [14]์—์„œ ์‚ฌ์šฉํ•œ ๋ฐ์ดํ„ฐ ์ฆ์‹์„ ์ด์šฉํ•˜์˜€๋‹ค. ํžˆ์Šคํ† ๊ทธ๋žจ ํ‰ํ™œํ™”๋ฅผ ์‚ฌ์šฉํ•œ ๋ฐฉ์‹์€ ์ถ”์ถœ๋œ ์„ ํ™” ๋ฐ์ดํ„ฐ์˜ ๋ถ„ํฌ๋ฅผ ์กฐ์ •ํ•˜์—ฌ ํ•™์Šต ํ›„ ์‹ค์ œ ์„ ํ™”์— ์„œ ์•ˆ์ •์ ์ธ ์ฑ„์ƒ‰์ด ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜์˜€๋‹ค. ํ•™์Šต ๊ณผ์ •์—์„œ ์„ ํ™” ์ถ”์ถœ๋ฐฉ์‹์„ XDoG (Extended Difference of Gaussians) [17] ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ Dilate abs sub [14]์„ ๊ฐ™์ด ์‚ฌ์šฉํ•ด ํ•™์Šต ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๊ณผ์ ํ•ฉ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜์˜€๋‹ค.

3.2 ์ด์ค‘ ์ƒ์„ฑ์ž

Lee et al. [14]์—์„œ ์‚ฌ์šฉํ•œ ์ด์ค‘ ์ƒ์„ฑ์ž ๋ฐฉ์‹์€ ์ฑ„์ƒ‰์˜ ์ „์ฒด ๊ณผ์ •์„ ๋‘ ๋‹จ๊ณ„๋กœ ๋‚˜๋ˆ„์–ด ๋จผ์ € ์ดˆ์•ˆ์„ ์ƒ์„ฑํ•˜๊ณ , ๋‘ ๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ ์ƒ์„ฑ๋œ ์ดˆ์•ˆ์„ ์‚ฌ์šฉํ•ด ๋” ํฐ ์„ ํ™”๋ฅผ ์ฑ„์ƒ‰ํ•œ๋‹ค. ์ด์ค‘ ์ƒ์„ฑ์ž ๋ฐฉ์‹์€ ์ € ํ•ด์ƒ๋„์—์„œ ๊ณ ํ•ด์ƒ๋„๋กœ ์ ์ง„์ ์ธ ํ•™์Šต์„ ์‚ฌ์šฉํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์•ˆ์ •์  ์ธ ํ•™์Šต์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๋ฉฐ ๊ฐ ๋‹จ๊ณ„์— ์ ํ•ฉํ•œ ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋‹ค. ํ’๋ถ€ํ•œ ์ƒ‰์ƒ์„ ์ƒ์„ฑํ•  ํ•„์š”๊ฐ€ ์žˆ์œผ๋ฏ€๋กœ ์ดˆ์•ˆ(Draft) ๋ชจ๋ธ์€ GAN์„ ์‚ฌ์šฉํ•ด ์›๋ณธ ์ปฌ๋Ÿฌ ์ด๋ฏธ์ง€์™€ ์œ ์‚ฌํ•œ ๋ถ„ํฌ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋„๋ก ํ•™์Šตํ•œ๋‹ค. ์ดˆ์•ˆ ๋‹จ๊ณ„์—์„œ๋Š” ๊ณ ํ’ˆ์งˆ์˜ ์ด๋ฏธ์ง€ ์ƒ์„ฑ์ด ํ•„์š”ํ•˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์— ๋‹ค์–‘ํ•œ ์ปฌ๋Ÿฌ๋ฅผ ๊ฐ€์ง€๋Š” ์ €ํ•ด์ƒ๋„(128 x 128 pixel) ์ด๋ฏธ์ง€๋ฅผ ์ƒ์„ฑํ•˜์—ฌ ํ•™์Šต ๋ณต์žก๋„๋ฅผ ์ค„์˜€๋‹ค. ์ฑ„์ƒ‰(Colorization) ๋ชจ๋ธ ๋‹จ๊ณ„์—์„œ๋Š” ์ดˆ์•ˆ ๋ชจ๋ธ์—์„œ ๋งŒ๋“ค์–ด์ง„ ์ด๋ฏธ์ง€๋ฅผ ์‚ฌ์šฉํ•ด ๊ณ ํ•ด์ƒ๋„(512 x 512 pixel)์˜ ์„ ํ™”๋ฅผ ์ฑ„์ƒ‰ํ•˜๋Š” ์ž‘์—…์„ ์ง„ํ–‰ํ•œ๋‹ค. ์ดˆ์•ˆ ์ด๋ฏธ์ง€์—์„œ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ๋Š” ์ƒ‰ ๋ฒˆ์ง, ์™œ๊ณก๊ณผ ๊ฐ™์€ ์ธ๊ณต๋ฌผ์„ ์ œ๊ฑฐํ•˜๊ธฐ ์œ„ํ•ด ์ฑ„์ƒ‰ ๋‹จ๊ณ„์˜ ํ•™์Šต ๊ณผ์ •์—์„œ ์ธ๊ณต๋ฌผ์„ ํ•ฉ์„ฑํ•˜๋Š” ๊ณผ์ •์„ ํฌํ•จํ•ด ํ•™์Šตํ•œ๋‹ค. ์ดˆ์•ˆ ๋ฐ ์ฑ„์ƒ‰ ๋ชจ๋ธ ๋ชจ๋‘ ์ด๋ฏธ์ง€ ๋Œ€ ์ด๋ฏธ์ง€ ๋ณ€ํ™˜์— ์ฃผ๋กœ ์‚ฌ์šฉ๋˜๋Š” U-Net [18] ๊ตฌ์กฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ตฌ์„ฑํ•˜์˜€๋‹ค [14,18].

3.3 ์„ ํ™” ์†์‹ค ํ•จ์ˆ˜

์„ ํ™”์— ๋Œ€ํ•œ ๊ณผ์ ํ•ฉ์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ ์ถ”๊ฐ€์ ์ธ ๋ฐฉ๋ฒ•์œผ๋กœ ์„ ํ™” ํƒ์ง€ ๋ชจ๋ธ(LDM, Line Detection Model)์„ ์‚ฌ์šฉํ•œ ์„ ํ™” ์†์‹ค Lline ์ด ์žˆ๋‹ค. LDM์€ ์˜คํ† ์ธ์ฝ”๋” ๊ตฌ์กฐ๋กœ ์ปฌ๋Ÿฌ ์ด๋ฏธ์ง€์—์„œ ์„ ํ™” ์ด๋ฏธ์ง€๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ์ž‘์—…์„ ์ง„ํ–‰ํ•œ๋‹ค. ์‚ฌ์ „ ํ•™์Šต๋œ LDM์„ ์ดˆ์•ˆ ๋ชจ๋ธ ํ•™์Šต ๋‹จ๊ณ„์— ์‚ฌ์šฉํ•ด ์ƒ์„ฑ๋œ ์ดˆ์•ˆ ์ด๋ฏธ์ง€์—์„œ ์„ ํ™”๋ฅผ ์ถ”์ถœํ•œ๋‹ค. ์ดํ›„ ์ถ”์ถœ๋œ ์„ ํ™”์™€ ์ดˆ์•ˆ ๋ชจ๋ธ์— ์ž…๋ ฅ๋˜๋Š” ์„ ํ™”์˜ L1 ์ฐจ์ด๋ฅผ ๊ณ„์‚ฐํ•œ๋‹ค. ์„ ํ™” ์†์‹ค์€ LDM์— ์˜ํ•ด ์ƒ์„ฑ๋œ ๋‘ ์„ ํ™”์˜ ์ฐจ์ด๋ฅผ ์ค„์ž„์œผ๋กœ ์ดˆ์•ˆ ๋ชจ๋ธ์—์„œ ์„ ํ™” ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์„ ๋†’์ธ๋‹ค. ์„ ํ™” ์†์‹ค Lline ์€ ์ˆ˜์‹ 1 ๋‚˜ํƒ€๋‚ด๋ฉฐ ์ˆ˜์‹์—์„œ G๋Š” ์ดˆ์•ˆ ์ƒ์„ฑ์ž, ldm์€ ์„ ํ™” ํƒ์ง€ ๋ชจ๋ธ, l ,h ,c๋Š” ๊ฐ๊ฐ ์„ ํ™”(128 x 128 pixel), ํžŒํŠธ, ์ปฌ๋Ÿฌ ์ด๋ฏธ์ง€๋ฅผ ๋‚˜ํƒ€๋‚ธ๋‹ค.

์ˆ˜์‹ 1
$$ \mathcal{L}_{l i n e}(G, l d m)=\mathbb{E}_{l, h, c}\left[\|l d m(c)-l d m(G(l, h))\|_{1}\right] $$

3.4 ์ฃผํŒŒ์ˆ˜ ๋ถ„ํ• 

์„œ๋น„์Šค์—์„œ ์ถœ๋ ฅ ํ•ด์ƒ๋„๋ฅผ ๋†’์ด๊ธฐ ์œ„ํ•ด Lee et al. [19]์—์„œ ์ œ์•ˆํ•œ ์ฃผํŒŒ์ˆ˜ ๋ถ„ํ• ์„ ์‚ฌ์šฉํ–ˆ๋‹ค. ์ž…๋ ฅ ์„ ํ™”์—์„œ ๊ณ ํ•ด์ƒ๋„์— ํ•„์š”ํ•œ ๊ณ ์ฃผํŒŒ ์„ฑ๋ถ„์„ ์ƒ์„ฑํ•˜์˜€๊ณ , ์ฑ„์ƒ‰ ๋ชจ๋ธ์—์„œ ์ƒ์„ฑ๋œ ์ฑ„์ƒ‰ ์ด๋ฏธ์ง€์—์„œ ์ €์ฃผํŒŒ ์„ฑ๋ถ„์„ ์ถ”์ถœํ•ด ํ•ฉ์„ฑํ•˜๋Š” ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•œ๋‹ค. ํ•ฉ์„ฑ ๊ณผ์ •์—์„œ๋Š” Linear light ํ˜ผํ•ฉ ๋ชจ๋“œ๋ฅผ ์‚ฌ์šฉํ•˜์˜€๋‹ค. ํ˜ผํ•ฉ ์‹œ ๋ฐ๊ธฐ ๊ฐ’ 127 (50\% ํšŒ์ƒ‰)์„ ๊ธฐ์ค€์œผ๋กœ, ๊ธฐ์ค€ ์ด์ƒ์˜ ๊ฐ’์€ ๋ฐ๊ฒŒ ๊ธฐ์ค€ ์ดํ•˜์˜ ๊ฐ’์€ ์–ด๋‘ก๊ฒŒ ํ•ฉ์„ฑํ•ด ์ž์—ฐ์Šค๋Ÿฌ์šด ํ•ฉ์„ฑ ๊ฒฐ๊ณผ๋ฌผ์„ ๋งŒ๋“ค์—ˆ๋‹ค. ์ข€ ๋” ์ƒ์„ธํ•œ ์ฃผํŒŒ์ˆ˜ ๋ถ„ํ• ์˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ๋Œ€ํ•œ ์„ค๋ช…์€ ์—ฐ๊ตฌ [19]์— ์†Œ๊ฐœ๋˜์—ˆ๋‹ค.

4. ํ”Œ๋žซํผ ์†Œ๊ฐœ

์ด ์žฅ์—์„œ๋Š” ์ œ์•ˆํ•˜๋Š” ์ฑ„์ƒ‰ ์„œ๋น„์Šค์™€ ๊ตฌ์กฐ๋ฅผ ์„ค๋ช…ํ•œ๋‹ค. ์„œ๋น„์Šค๋Š” ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ, ๋ชจ๋ธ ๊ด€๋ฆฌ ๋ฐ ์ถ”๋ก ์„ ์œ„ํ•œ ์ถ”๋ก  ์„œ๋ฒ„์™€ ์‚ฌ์šฉ์ž๋ฅผ ์œ„ํ•œ ์„œ๋น„์Šค ํ”„๋ŸฐํŠธ ์—”๋“œ๋กœ ๊ตฌ์„ฑ๋œ๋‹ค. ์ถ”๋ก  ์„œ๋ฒ„๋Š” ๊ณ ์ฐจํ•จ์ˆ˜ ์กฐํ•ฉ์„ ์‚ฌ์šฉํ•ด ๋ฐ์ดํ„ฐ๋ฅผ ์ „์ฒ˜๋ฆฌํ•˜๋Š” ์ „์ฒ˜๋ฆฌ๊ธฐ์™€ ์‹ฑ๊ธ€ ํ†ค ํŒจํ„ด์œผ๋กœ ์ถ”๋ก  ๋ชจ๋ธ์„ ๊ด€๋ฆฌํ•˜๋Š” ๋ชจ๋ธ ๊ด€๋ฆฌ์ž๋กœ ๊ตฌ์„ฑ๋œ๋‹ค. ์„œ๋น„์Šค๋Š” ์ถ”๋ก  ์„œ๋ฒ„์™€ ์„œ๋น„์Šค ํ”„๋ŸฐํŠธ ์—”๋“œ๋ฅผ ๊ฐ€์ƒํ™” ๊ธฐ๋ฒ•์ธ Docker๋ฅผ ์‚ฌ์šฉํ•ด ์ด๋ฏธ์ง€๋กœ ๋งŒ๋“ค์–ด ๊ตฌ์„ฑํ•œ๋‹ค. ์ œ์•ˆํ•˜๋Š” ์‹œ์Šคํ…œ ๋‹ค์ด์–ด๊ทธ๋žจ์€ ๊ทธ๋ฆผ 2์™€ ๊ฐ™๋‹ค.

๊ทธ๋ฆผ 2. ์‹œ์Šคํ…œ ๋‹ค์ด์–ด๊ทธ๋žจ

Fig. 2. System Diagram

../../Resources/kiee/KIEEP.2022.71.1.41/fig2.png

4.1 ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ ๋ชจ๋“ˆ

์ฑ„์ƒ‰ ๋ชจ๋ธ์€ ์„ ํ™”๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๋‹ค์–‘ํ•œ ์ž…๋ ฅ์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์ „์ฒ˜๋ฆฌ๋ฅผ ์œ„ํ•œ ์‹œ์Šคํ…œ์œผ๋กœ ํŒŒ์ด์ฌ์˜ ํด๋ž˜์Šค ํ˜ธ์ถœ ๋ฉ”์†Œ๋“œ๋ฅผ ์‚ฌ์šฉํ•ด ๊ณ ์ฐจํ•จ์ˆ˜์˜ ์กฐํ•ฉ์œผ๋กœ ๊ตฌ์„ฑํ•œ๋‹ค. ์ œ์•ˆํ•˜๋Š” ์‹œ์Šคํ…œ์€ ์ •๊ทœํ™”, ํฌ๊ธฐ์กฐ์ •, ํ…์„œ ๋ณ€ํ™˜, Lambda ํ•จ์ˆ˜ ๊ตฌํ˜„์„ ์œ„ํ•œ ๊ณ ์ฐจํ•จ์ˆ˜๋ฅผ ๋งŒ๋“ค์–ด ์‚ฌ์šฉํ•œ๋‹ค. ์ž…๋ ฅ์„ ์œ„ํ•œ ์„ ํ™” ๋ฐ์ดํ„ฐ๋กœ ํด๋ผ์ด์–ธํŠธ์—์„œ ๋ฐ›์€ ์„ ํ™” ์ด๋ฏธ์ง€๋ฅผ ์ดˆ์•ˆ ๋ชจ๋ธ ๋ฐ ์ฑ„์ƒ‰ ๋ชจ๋ธ์— ๋”ฐ๋ผ ํฌ๊ธฐ๋ฅผ ์กฐ์ •ํ•˜์—ฌ ์‚ฌ์šฉํ•˜์˜€๋‹ค. ํžŒํŠธ๋Š” ํด๋ผ์ด์–ธํŠธ๊ฐ€ ์ž…๋ ฅ๋ฐ›์€ ํˆฌ๋ช…๋„ ๋ฐ์ดํ„ฐ๊ฐ€ ๋ณด์กด๋œ PNG ๋ฐ์ดํ„ฐ์—์„œ ํžŒํŠธ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด RGB์™€ alpha ์ฑ„๋„์„ ๋ถ„๋ฆฌํ•ด ๋”ฐ๋กœ ์ „ ์ฒ˜๋ฆฌํ•˜์˜€๋‹ค.

4.2 ์ถ”๋ก  ๋ชจ๋ธ ๊ด€๋ฆฌ์ž

์ฑ„์ƒ‰ ๋ชจ๋ธ์€ ํŠน์„ฑ์— ๋”ฐ๋ผ ๋‹ค๋ฅธ ์ „ํ›„ ์ฒ˜๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•œ๋‹ค. ๋”ฐ๋ผ์„œ ๋ชจ๋“  ์ฑ„์ƒ‰ ๋ชจ๋ธ์€ ์ถ”์ƒ ํด๋ž˜์Šค๋ฅผ ์ƒ์†๋ฐ›์•„ ์‹ฑ๊ธ€ํ†ค ํŒจํ„ด์œผ๋กœ ๋ชจ๋ธ์„ ๊ด€๋ฆฌํ•œ๋‹ค. ๊ฐ ๋ชจ๋ธ์„ ์œ„ํ•ด ์ „/ํ›„ ์ฒ˜๋ฆฌ ๋ฐ ์ถ”๋ก  ๋ฉ”์†Œ๋“œ๋ฅผ ๊ตฌํ˜„ํ•˜์˜€๋‹ค. ์ถ”๋ก  ๋ชจ๋ธ ๊ด€๋ฆฌ์ž๋Š” ์„œ๋ฒ„๊ฐ€ ์‹œ์ž‘ํ•˜๊ธฐ ์ „ ์ธ์Šคํ„ด์Šคํ™”ํ•˜์—ฌ ์‚ฌ์šฉํ•œ๋‹ค. Pytorch๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์€ ONNX ๋ชจ๋ธ๋กœ ๋ณ€ํ™˜ํ•˜๋ฉฐ ์ถ”๋ก  ๊ณผ์ •์—์„œ๋Š” ONNX ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์‚ฌ์šฉํ•ด CPU์—์„œ ์ถ”๋ก ์„ ์ง„ํ–‰ํ•œ๋‹ค. ONNX๋Š” ์ถ”๋ก ์„ ์œ„ํ•œ ๋Ÿฐํƒ€์ž„์ด ๊ฐ€๋ณ๊ณ  ๋‹ค์–‘ํ•œ ํ”Œ๋žซํผ์—์„œ ๊ฐ€์†์„ ์ง€์›ํ•œ๋‹ค๋Š” ์žฅ์ ์ด ์žˆ๋‹ค. ์„œ๋น™์€ ๋งˆ์ดํฌ๋กœ์„œ๋น„์Šค ์•„ํ‚คํ…์ฒ˜ ํ”„๋ ˆ์ž„์›Œํฌ์ธ Flask๋ฅผ ์‚ฌ์šฉํ•ด ์„œ๋ฒ„๋ฅผ ๊ตฌ์„ฑํ•œ๋‹ค. ๊ฐœ๋ฐœ๋œ ์„œ๋น„์Šค๋Š” ์›น์„ ํ†ตํ•ด โ€œomnissiah.ys2lee.comโ€์—์„œ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋‹ค.

4.3 ์„œ๋น„์Šค ํ”„๋ŸฐํŠธ ์—”๋“œ

์„œ๋น„์Šค ํ”„๋ŸฐํŠธ ์—”๋“œ์˜ ๊ตฌ์กฐ๋Š” ๊ทธ๋ฆผ 3๊ณผ ๊ฐ™๊ณ  ๊ฐœ๋ฐœ๋œ ์„œ๋น„์Šค ํ˜•์ƒ์€ ๊ทธ๋ฆผ 1๋กœ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค. ์ž๋™์ฑ„์ƒ‰ ์„œ๋น„์Šค๋ฅผ ์œ„ํ•œ UI๋Š” React ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•ด ์›น ์„œ๋น„์Šค๋กœ ์ œ์ž‘ํ•œ๋‹ค. ์„œ๋น„์Šค๋Š” ์„ ํ™”๋ฅผ ๋ถˆ๋Ÿฌ์˜ค๊ณ  ํžŒํŠธ ์ •๋ณด๋ฅผ ์ปจํŠธ๋กค ํ•˜๋ฉฐ ์„ ํ™” ๋ฐ ํžŒํŠธ ๊ทธ๋ฆฌ๊ณ  ์„œ๋ฒ„๋ฅผ ํ†ตํ•ด ์ฑ„์ƒ‰๋œ ์ฑ„์ƒ‰ ์ด๋ฏธ์ง€๋ฅผ ์‹œ๊ฐํ™”ํ•œ๋‹ค. ์„œ๋น„์Šค ํ”„๋ŸฐํŠธ ์—”๋“œ๋Š” โ€œPaintsToolsโ€, โ€œPaintsCanvasโ€, โ€œPaintsViewerโ€์˜ ์„ธ ๊ฐœ์˜ ์ฃผ์š” ์ปดํฌ๋„ŒํŠธ๋กœ ๊ตฌ์„ฑ๋œ๋‹ค. โ€œPaintsToolsโ€๋Š” ํžŒํŠธ ์ƒ‰ ๋ณ€๊ฒฝ, ํžŒํŠธ ์‚ฌ์ด์ฆˆ ๋ณ€๊ฒฝ, ํžŒํŠธ ์ƒํƒœ ์ปจํŠธ๋กค ๋ฐ ์ฑ„์ƒ‰ ์ง„ํ–‰ ๊ทธ๋ฆฌ๊ณ  ์ €์žฅ๊ณผ ๊ฐ™์€ ์ปจํŠธ๋กค ๊ธฐ๋Šฅ ์„ ๋‹ด๋‹นํ•œ๋‹ค. โ€œPaintsCanvasโ€๋Š” ์„ ํ™” ๋ฐ ํžŒํŠธ๋ฅผ ์‹œ๊ฐํ™”ํ•˜๊ณ  ๋ ˆ์ด์–ด ๊ตฌ์กฐ๋กœ ์„ ํ™”์— ํžŒํŠธ๋ฅผ ์ž…๋ ฅํ•˜๋Š” ๊ธฐ๋Šฅ์„ ๋‹ด๋‹นํ•œ๋‹ค. โ€œPaintsViewerโ€๋Š” ์„œ๋ฒ„์—์„œ ์ฑ„์ƒ‰๋œ ์ด๋ฏธ์ง€๋ฅผ ์‹œ๊ฐํ™”ํ•˜๋Š” ๊ธฐ๋Šฅ์„ ๋‹ด๋‹นํ•œ๋‹ค.

๊ทธ๋ฆผ 3. ์„œ๋น„์Šค ํ”„๋ŸฐํŠธ ์—”๋“œ ์ปดํฌ๋„ŒํŠธ ํŠธ๋ฆฌ

Fig. 3. Service Frontend Component Tree

../../Resources/kiee/KIEEP.2022.71.1.41/fig3.png

5. ์‹คํ—˜๊ณผ ๋ถ„์„

5.1 ์‹คํ—˜ ํ™˜๊ฒฝ

์ œ์•ˆํ•œ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ ๊ตฌํ˜„ ๋ฐ ํ•™์Šต์„ ์œ„ํ•ด PyTorch ํ”„๋ ˆ์ž„์›Œํฌ [20]๋ฅผ ์‚ฌ์šฉํ–ˆ๋‹ค. ํ•™์Šต์„ ์œ„ํ•œ ๋ฐ์ดํ„ฐ ์…‹์€ shuushuu-imageboard [21]์—์„œ ์ˆ˜์ง‘ํ•˜์˜€๋‹ค. ๋ฐ์ดํ„ฐ๋Š” ํ•™์Šต์— ๋ถ€์ •์ ์ธ ์˜ํ–ฅ์„ ์ค„ ์ˆ˜ ์žˆ๋Š” ๋ฐ์ดํ„ฐ๋ฅผ ํ•„ํ„ฐ๋งํ•˜์—ฌ ์•ฝ 70๋งŒ ์žฅ์œผ๋กœ ๊ตฌ์„ฑํ–ˆ๋‹ค. ํ•„ํ„ฐ๋ง ๋œ ๋ฐ์ดํ„ฐ๋Š” ํ‘๋ฐฑ, ํ•˜์ด/๋กœ์šฐ ํ‚ค ์ด๋ฏธ์ง€, 512 pixel ์ดํ•˜์˜ ์ž‘์€ ์ด๋ฏธ์ง€, ์ „๋ฐ˜์ ์ธ ํ†ค ํ˜น์€ ์ƒ‰๊ฐ์ด ํ•œ์ชฝ์œผ๋กœ ํŽธํ–ฅ๋œ ์ด๋ฏธ์ง€ ๋“ฑ์ด๋‹ค. ๋ชจ๋ธ ํ•™์Šต ์„ ์œ„ํ•ด์„œ ํ•œ ์žฅ์˜ NVIDIA RTX 3090์„ ์‚ฌ์šฉ ํ–ˆ์œผ๋ฉฐ Arch Linux ํ™˜๊ฒฝ์—์„œ ํ•™์Šต์„ ์ง„ํ–‰ํ–ˆ๋‹ค. ์„œ๋น„์Šค๋ฅผ ๊ตฌ๋™ํ•˜๋Š” ์‹คํ—˜ ํ™˜๊ฒฝ์€ Intel i9-9980HZ, 32GB RAM์˜ MacOS์—์„œ ํ…Œ์ŠคํŠธ๋ฅผ ์ง„ํ–‰ํ–ˆ๋‹ค. ์ƒ์„ธ ๊ตฌ์„ฑ์€ ํ‘œ 1์— ๋‚˜ํƒ€๋‚ด์—ˆ๋‹ค.

๊ทธ๋ฆผ 4. ์‹œ๊ฐ์  ๋น„๊ต (a): Tanpopo w/o hint [2], (b): Satsuki w/o hint [2], (c): Canna w/o hint [2], (d): Ours w/o hint [14], (e): Tanpopo [2], (f): Satsuki [2], (g): Canna [2], (h): Ours [14]

Fig. 4. Visual Comparison (a): Tanpopo w/o hint [2], (b): Satsuki w/o hint [2], (c): Canna w/o hint [2], (d): Ours w/o hint [14], (e): Tanpopo [2], (f): Satsuki [2], (g): Canna [2], (h): Ours [14]

../../Resources/kiee/KIEEP.2022.71.1.41/fig4.png

ํ‘œ 1 ์‹คํ—˜ ํ™˜๊ฒฝ

Table 1 Test Environment

HW

Specification

SW

Version

Training

CPU

Intel i9-10980XE

Python

3.8.5

GPU

NVIDIA RTX 3090

Pytorch

1.9

RAM

128GB

ONNX

1.10.2

OS

Arch Linux

Serving

CPU

Intel i9-9980HZ

Python

3.8.5

RAM

32GB

onnxruntime

1.8.1

OS

MacOS

numpy

1.21.2

flask

2.0.1

5.2 ์ถ”๋ก  ํ™˜๊ฒฝ ์„ฑ๋Šฅ ๋น„๊ต

์ œ์•ˆํ•œ ์„œ๋น„์Šค๋Š” CPU์—์„œ ๋ชจ๋ธ ์ถ”๋ก ์„ ์‚ฌ์šฉํ•œ๋‹ค. ์‹คํ—˜์€ CPU ์—์„œ Pytorch์˜ TorchScript ๋ชจ๋ธ๊ณผ onnxruntime์„ ์‚ฌ์šฉํ•ด 512 x 512 pixel ํ•ด์ƒ๋„ ์ด๋ฏธ์ง€์˜ ์ถ”๋ก  ์‹œ๊ฐ„์„ ๋น„๊ตํ–ˆ๋‹ค. ์‹คํ—˜์€ ๊ฐ๊ฐ 100ํšŒ์˜ ์ถ”๋ก ์„ ์ง„ํ–‰ํ•ด ํ‰๊ท  ์‹œ๊ฐ„์„ ๋น„๊ตํ•˜์˜€๋‹ค. ์‹คํ—˜์— ์‚ฌ์šฉํ•œ ์ž…๋ ฅ์€ ๋‘ ๋ชจ๋ธ ๋ชจ๋‘ ์„ ํ™” 1x1x512x512, 1x1x128x128 ๊ทธ๋ฆฌ๊ณ  ํžŒํŠธ 1x4x128x128 ์‚ฌ์ด์ฆˆ์˜ 0์œผ๋กœ ์ด๋ฃจ์–ด์ง„ ๋ฒกํ„ฐ๋ฅผ ์‚ฌ์šฉํ•ด ์ง„ํ–‰ํ•œ๋‹ค. ํ™˜๊ฒฝ์— ๋”ฐ๋ฅธ ์ถ”๋ก  ์†๋„ ๋น„๊ต ๊ฒฐ๊ณผ๋ฅผ ํ‘œ 2์— ๋‚˜ํƒ€๋‚ด์—ˆ๋‹ค. ONNX๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐœ๋ฐœํ•œ ์‹œ์Šคํ…œ์˜ ์‹คํ–‰ ํ‰๊ท ๊ฐ’์ด 0.4040 ์ดˆ๋กœ Torchscript์™€ ๋น„๊ตํ•ด 5๋ฐฐ ์ด์ƒ์˜ ์†๋„๋ฅผ ๋ณด์ด๋ฉฐ ํšจ์œจ์ ์ธ ์ถ”๋ก ์„ ํ•˜๋Š” ๊ฒƒ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค.

5.3 ์ฑ„์ƒ‰ ์„ฑ๋Šฅ ์‹œ๊ฐ์  ๋น„๊ต

์ œ์•ˆํ•œ ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•œ ์„œ๋น„์Šค์™€ Petalica Paint [2]์˜ โ€œTanpopoโ€, โ€œSatsukiโ€, โ€œCannaโ€ ๊ฐ๊ฐ์˜ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•ด ์‹ค์ œ ์„ ํ™”์˜ ์ฑ„์ƒ‰์„ฑ๋Šฅ์„ ์‹œ๊ฐ์ ์œผ๋กœ ๋น„๊ตํ•˜์˜€๋‹ค. ์ƒ์šฉ ์„œ๋น„์Šค์˜ ๊ฒฝ์šฐ ์†Œ์Šค ์ฝ”๋“œ๊ฐ€ ๊ณต๊ฐœ๋˜์ง€ ์•Š์•„ ๊ฐ™์€ ํžŒํŠธ๋ฅผ ์‚ฌ์šฉํ•œ ๋น„๊ต๊ฐ€ ๋ถˆ๊ฐ€๋Šฅํ•˜๋ฏ€๋กœ ํžŒํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š์€ ์ฑ„์ƒ‰ ๊ฒฐ๊ณผ๋ฅผ ๋น„๊ตํ•œ๋‹ค. ์ฑ„์ƒ‰ ๊ฒฐ๊ณผ๋Š” ๊ทธ๋ฆผ 4์— ๋‚˜ํƒ€๋‚ด์—ˆ๋‹ค. ๊ทธ๋ฆผ 4(a-d) ์ด๋ฏธ์ง€๋ฅผ ํ™•์ธํ•˜๋ฉด ์„ ํ™” ํŠน์„ฑ์— ๋”ฐ๋ผ ํžŒํŠธ๊ฐ€ ์—†์„ ๋•Œ ์ฑ„์ƒ‰์ด ์–ด๋–ค ์‹์œผ๋กœ ์ง„ํ–‰๋˜๋Š”์ง€ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค.

ํ‘œ 2 ์ถ”๋ก  ์„ฑ๋Šฅ

Table 2 Inference Performance

Runtime

Mean(Sec)

Std(Sec)

ONNX

0.4040

0.0127

TorchScript

2.2683

0.0754

์‹คํ—˜ ๊ฒฐ๊ณผ ๊ทธ๋ฆผ 4(a, e)๋Š” checkerboard artifacts ํ˜„์ƒ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ๊ณ , ๊ทธ๋ฆผ 4(c, g)์˜ โ€œCannaโ€ ๋ชจ๋ธ์€ ์ƒ‰์ƒ์˜ ์™œ๊ณก์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค. ๊ทธ๋ฆผ 4(d)์˜ 2๋ฒˆ์งธ ์ด๋ฏธ์ง€๋ฅผ ๋ณด๋ฉด ์ œ์•ˆํ•˜๋Š” ๋ชจ๋ธ์—์„œ๋„ ํŠน์ • ์„ ํ™”์— ๋”ฐ๋ผ ์ฑ„์ƒ‰์ด ๋ถˆ์•ˆ์ •ํ•œ ๊ฒƒ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์ง€๋งŒ ํžŒํŠธ๋ฅผ ์‚ฌ์šฉ ํ•œ 4(h)๋Š” ์•ˆ์ •์ ์œผ๋กœ ์ฑ„์ƒ‰ํ•œ ๊ฒƒ์„ ํ™•์ธํ–ˆ๋‹ค.

6. ๊ฒฐ ๋ก 

๋ณธ ์—ฐ๊ตฌ๋Š” ์ž๋™์ฑ„์ƒ‰ ๋ถ„์•ผ์—์„œ ๊ฐœ๋ฐœ๋œ ๋ชจ๋ธ์„ ์„œ๋น™ํ•˜๊ธฐ ์œ„ํ•œ ํ”Œ๋žซํผ์ด ์—†๋Š” ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ONNX๋ฅผ ์‚ฌ์šฉํ•œ ์ž๋™์ฑ„์ƒ‰ ๋ชจ๋ธ ์„œ๋น„์Šค๋ฅผ ๊ฐœ๋ฐœํ–ˆ๋‹ค. ์„œ๋น„์Šค๋Š” ๊ณ ์ฐจํ•จ์ˆ˜๋ฅผ ํ™œ์šฉํ•œ ์ „์ฒ˜๋ฆฌ ๊ธฐ๋ฒ•์„ ์ œ์•ˆํ•˜๊ณ  ONNX๋ฅผ ์‚ฌ์šฉํ•ด CPU ํ™˜๊ฒฝ์—์„œ๋„ ํšจ์œจ์ ์ธ ์ถ”๋ก ์ด ๊ฐ€๋Šฅํ•˜๋„๋ก ๊ฐœ๋ฐœํ•˜์˜€๋‹ค. ์ฑ„์ƒ‰์„ ์œ„ํ•œ ๋ชจ๋ธ์€ ํ•™์Šต ์„ ํ™” ๋ฐ์ดํ„ฐ์˜ ๊ณผ์ ํ•ฉ ํ˜„์ƒ์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ์„ ํ™” ๋ฐ์ดํ„ฐ ์ฆ์‹์„ ์‚ฌ์šฉํ•˜๊ณ , ํ•™์Šต ๊ณผ์ •์— ์„œ ์„ ํ™” ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด ์„ ํ™” ์†์‹ค ํ•จ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•˜์˜€๋‹ค. ์ฑ„์ƒ‰ ๋ชจ๋ธ์€ ํ•™์Šต ์•ˆ์ •์„ฑ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด ์ดˆ์•ˆ ์ƒ์„ฑ๊ณผ ์ฑ„์ƒ‰ ์ž‘์—…์„ ๋ถ„๋ฆฌํ•œ ์ด์ค‘ ์ƒ์„ฑ์ž๋กœ ๋ชจ๋ธ์„ ๊ตฌ์„ฑํ•œ๋‹ค. ์„œ๋น„์Šค์—์„œ ์ถœ๋ ฅ ํ•ด์ƒ๋„๋ฅผ ๋†’์ด๊ธฐ ์œ„ํ•ด ์ฃผํŒŒ์ˆ˜ ๋ถ„ํ•  ๊ธฐ๋ฒ•์„ ํ™œ์šฉํ•œ ์ด๋ฏธ์ง€ ํ•ฉ์„ฑ์„ ์‚ฌ์šฉํ•˜์˜€๋‹ค. ์„œ๋น„์Šค์˜ ์„ฑ๋Šฅ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด CPU์—์„œ torchscript ์™€ onnxrumtime ์„ ์‚ฌ์šฉํ•ด ์ถ”๋ก  ์„ฑ๋Šฅ์„ ๋น„๊ตํ•œ ๊ฒฐ๊ณผ ONNX๋กœ ๊ฐœ๋ฐœํ•œ ์‹œ์Šคํ…œ์ด ํ‰ ๊ท  0.4040์ดˆ์˜ ์ถ”๋ก ์„ฑ๋Šฅ์„ ๋ณด์—ฌ torchscript์˜ ์„ฑ๋Šฅ๊ณผ ๋น„๊ตํ•ด 5๋ฐฐ ์ด์ƒ์˜ ๋น ๋ฅธ ์ถ”๋ก ์ด ๊ฐ€๋Šฅํ•œ ๊ฒƒ์„ ํ™•์ธํ–ˆ๋‹ค. ๊ธฐ์กด ์„œ๋น„์Šค์™€ ์ฑ„์ƒ‰ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•˜๊ธฐ ์œ„ํ•ด Petalica Paint [2]์˜ 3๊ฐ€์ง€ ๋ชจ๋ธ โ€œTanpopoโ€, โ€œSatsukiโ€, โ€œCannaโ€์™€ ์‹œ๊ฐ์ ์œผ๋กœ ๋น„๊ตํ•˜์˜€๋‹ค. ๊ทธ ๊ฒฐ๊ณผ ํžŒํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š์€ ์ฑ„์ƒ‰์—์„œ๋Š” ๋ชจ๋“  ๋ชจ๋ธ์ด ์„ ํ™” ํŠน์ง•์— ๋”ฐ๋ผ ์ฑ„์ƒ‰ ์„ฑ๋Šฅ์ด ๋ถˆ์•ˆ์ •ํ–ˆ๋‹ค. ํžŒํŠธ๋ฅผ ์‚ฌ์šฉํ•  ๊ฒฝ์šฐ ๊ทธ๋ฆผ 4(h) ์ œ์•ˆํ•œ ์„œ๋น„์Šค์—์„œ ๊ธฐ์กด ์„œ๋น„์Šค ๊ทธ๋ฆผ 4(e)๋ณด๋‹ค ์ธ๊ณต๋ฌผ๊ณผ ์™œ๊ณก ์—†๋Š” ์šฐ์ˆ˜ํ•œ ํ’ˆ์งˆ์˜ ์ฑ„์ƒ‰ ์ด๋ฏธ์ง€๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์—ˆ๋‹ค.

Acknowledgements

This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. 2019R1G1A1100455).

References

1 
A. Radford, L. Metz, S. Chintala, 2015, Unsupervised representation learning with deep convolutional generative adversarial networks, arXiv preprint arXiv:1511.06434DOI
2 
pixiv inc., 2021, Petalica paint., https://petalica-paint.pixiv.dev/index_en.html[Online; accessed22 October โˆ’ 2021]URL
3 
P. Isola, J.-Y. Zhu, T. Zhou, A. A. Efros, 2017, Image-to-image translation with conditional adversarial networks, in Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 1125-1134URL
4 
S. Kang, J. Choo, J. Chang, 2017, Consistent comic colorization with pixel-wise background classification, in NIPSโ€™17 Workshop on Machine Learning for Creativity and DesignURL
5 
C. Furusawa, K. Hiroshiba, K. Ogaki, Y. Odagiri, 2017, Comicolorization: semi-automatic manga colorization, in SIGGRAPH Asia 2017 Technical Briefs, pp. 1-4DOI
6 
P. Hensman, K. Aizawa, 2017, cgan-based manga colorization using a single training image, in 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR), IEEE, Vol. 3, pp. 72-77DOI
7 
L. Zhang, Y. Ji, X. Lin, C. Liu, 2017, Style transfer for anime sketches with enhanced residual u-net and auxiliary classifier gan, in 2017 4th IAPR Asian Conference on Pattern Recognition (ACPR), IEEE, pp. 506-511DOI
8 
P. Sangkloy, J. Lu, C. Fang, F. Yu, J. Hays, 2017, Scribbler: Controlling deep image synthesis with sketch and color, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 5400-5409URL
9 
Y. Liu, Z. Qin, T. Wan, Z. Luo, 2018, Auto-painter: Cartoon image generation from sketch by using conditional wasserstein generative adversarial networks, Neurocomputing, Vol. 311, pp. 78-87DOI
10 
K. Frans, 2017, Outline colorization through tandem adversarial networks, arXiv preprint arXiv:1704.08834DOI
11 
Y. Ci, X. Ma, Z. Wang, H. Li, Z. Luo, 2018, User-guided deep anime line art colorization with conditional adversarial networks, in Proceedings of the 26th ACM international conference on Multimedia, pp. 1536-1544DOI
12 
L. Zhang, C. Li, T.-T. Wong, Y. Ji, C. Liu, 2018, Two-stage sketch colorization, ACM Transactions on Graphics (TOG), Vol. 37, No. 6, pp. 1-14Google Search
13 
Y. Hati, G. Jouet, F. Rousseaux, C. Duhart, 2019, Paintstorch: a user-guided anime line art colorization tool with double generator conditional adversarial network, in European Conference on Visual Media Production, pp. 1-10DOI
14 
Y. Lee, S. Lee, 2020, Automatic colorization of anime style illustrations using a two-stage generator, Applied Sciences, Vol. 10, No. 23, pp. 8699DOI
15 
The Linux Foundation, 2019, Onnx: Open neural network exchange., https://github.com/onnx/onnxURL
16 
K. Simonyan, A. Zisserman, Very deep convolutional networks for large-scale image recognition, in 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings (Y. Bengio and Y. LeCun, eds.), 2015.DOI
17 
H. Winnemรถller, J. E. Kyprianidis, S. C. Olsen, 2012, Xdog: an extended difference-of-gaussians compendium including advanced image stylization, Computers & Graphics, Vol. 36, No. 6, pp. 740-753DOI
18 
A. Odena, V. Dumoulin, C. Olah, 2016, Deconvolution and checkerboard artifacts, Distill, Vol. 1, No. 10, pp. e3URL
19 
Yeongseop Lee, Seongjin Lee, 2020, Automatic Colorization of High-resolution Animation Style Line-art based on Frequency Separation and Two-Stage Generator, The Transactions of the Korean Institute of Electrical Engineers, Vol. 69p, No. 4, pp. 275~283Google Search
20 
A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, 2019, Pytorch: An imperative style, high-performance deep learning library, Advances in neural information processing systems, Vol. 32URL
21 
E-Shuushuu, 2018, E-Shuushuu - Kawaii Image Board., https://e-shuushuu.netURL

์ €์ž์†Œ๊ฐœ

์ด์˜์„ญ(Yeongseop Lee)
../../Resources/kiee/KIEEP.2022.71.1.41/au1.png

Youngseop Lee graduated from Gyeongsang National University in 2020. He received master degree at the Dept of Information Science, Gyeongsang National University in 2022. After graduation, he joined Funzin. His research interests includes Machine Learning, Neural Network, Image Generation, and Image Processing.

์ด์„ฑ์ง„(Seongjin Lee)
../../Resources/kiee/KIEEP.2022.71.1.41/au2.png

Seongjin Lee graduated from Hanyang University in 2006. He recieved Master and Ph.D. degree in the same university in 2008 and 2015, respectively. He worked as postdoc in Storage Center Hanyang University till 2017 and became an assistant research professor there. He joined Gyeongsang National University in 2017 as an assistant professor. His research interest includes Operating System, Storage System, System Optimization, Avionics, and Machine Learning.