Mobile QR Code QR CODE

Journal of the Korea Concrete Institute

J Korea Inst. Struct. Maint. Insp.
  • Indexed by
  • Korea Citation Index (KCI)




๋”ฅ๋Ÿฌ๋‹, ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง, ์•„์ŠคํŒ”ํŠธ ๋„๋กœํฌ์žฅ, ์•„์ŠคํŒ”ํŠธ ๋„๋กœํฌ์žฅ ํ‘œ๋ฉด๊ท ์—ด
Deep learning, Convolutional Neural Network, Asphalt Pavement, Surface Crack

1. ์„œ ๋ก 

๊ตญ๋‚ด ์•„์ŠคํŒ”ํŠธ ์ฝ˜ํฌ๋ฆฌํŠธ ๋„๋กœํฌ์žฅ์— ๋Œ€ํ•œ ์œ ์ง€๊ด€๋ฆฌ๋Š” ์ž๋™ ํฌ์žฅ์ƒํƒœ ์กฐ์‚ฌ์žฅ๋น„์— ์˜ํ•ด ์ˆ˜์ง‘๋˜๋Š” ๋…ธ๋ฉด ์˜์ƒ์„ ๋ถ„์„ํ•˜๊ณ , ๊ฒฐํ•จ์ •๋„๋ฅผ ํŒŒ์•…ํ•˜์—ฌ ๋ณด์ˆ˜์‹œ๊ธฐ, ๋ณด์ˆ˜๋ฐฉ๋ฒ• ๋“ฑ์„ ๊ฒฐ์ •ํ•˜๊ฒŒ ๋œ ๋‹ค. ๋„๋กœํฌ์žฅ ํ‘œ๋ฉด๊ฒฐํ•จ์— ๋Œ€ํ•œ ๋ถ„์„์€ ์กฐ์‚ฌ๋œ ๋…ธ๋ฉด ์˜์ƒ์„ ํŠน ์ • ํฌ๊ธฐ์˜ ๊ฒฉ์ž ํ˜•ํƒœ๋กœ ๋ถ„ํ• ํ•˜์—ฌ ์œก์•ˆ์ ๊ฒ€์„ ํ†ตํ•ด ๊ฒฐํ•จ์„ ํฌ ํ•จํ•˜๋Š” ๊ฒฉ์ž๋ฅผ ํŒ๋ณ„ํ•˜๊ณ , ๊ฒฐํ•จ ๊ฒฉ์ž์ˆ˜๋ฅผ ์ „์ฒด ๊ฒฉ์ž์ˆ˜๋กœ ๋‚˜๋ˆˆ ๊ฐ’์„ ํ•ด๋‹น ๋…ธ๋ฉด ์˜์—ญ์˜ ๊ท ์—ด์œจ๋กœ ํŒ๋‹จํ•œ๋‹ค. ๊ตญ๋‚ด ์•„์ŠคํŒ”ํŠธ ์ฝ˜ ํฌ๋ฆฌํŠธ ๋„๋กœํฌ์žฅ ์ž๋™๋…ธ๋ฉด ์กฐ์‚ฌ์žฅ๋น„์˜ ์˜ˆ๋กœ KRISS(Korea Roadway Infrastructure Survey System)๋Š” 3.6m x 10m ์˜์—ญ์˜ ๋‹จ์œ„๋กœ ๋…ธ๋ฉด ์˜์ƒ์„ ์ˆ˜์ง‘ํ•˜๋ฉฐ, ์ด ๋…ธ๋ฉด ์˜์ƒ์„ ๊ฐ€๋กœ 300mm, ์„ธ๋กœ 300mm์˜ ๊ฒฉ์ž ํ˜•ํƒœ๋กœ ๋ถ„ํ• ํ•˜์—ฌ ๊ท ์—ด์œจ์„ ํ‰๊ฐ€ํ•œ๋‹ค (Kim et al., 2008).

ํฌ์žฅ ๊ฒฐํ•จ ๋ถ„์„์— ์žˆ์–ด์„œ ๊ฐ๊ฐ์˜ ๊ฒฉ์ž์— ๋Œ€ํ•œ ์œก์•ˆ์ ๊ฒ€์€ ๋งŽ์€ ๋น„์šฉ์„ ๋ฐœ์ƒ์‹œํ‚ค๊ณ  ๋”ฐ๋ผ์„œ ์ด๋ฅผ ์ž๋™ํ™”ํ•˜๋ ค๋Š” ์—ฐ๊ตฌ๊ฐ€ ์ˆ˜ํ–‰๋˜์–ด์™”๋‹ค(Koch and Brilakis, 2011; Sorncharean and Phiphobmongkol, 2008; Rababaah et al., 2005). ๊ธฐ์กด์˜ ์˜์ƒ ์ฒ˜๋ฆฌ๋ฅผ ํ†ตํ•œ ์•„์ŠคํŒ”ํŠธ ์ฝ˜ํฌ๋ฆฌํŠธ ๋„๋กœํฌ์žฅ์˜ ํ‘œ๋ฉด๊ฒฐํ•จ ๊ฒ€์‚ฌ๋Š” ์˜์ƒ์—์„œ ํŠน์ • ๊ฒฐํ•จ์˜ ํŠน์ง•์„ ์ถ”์ถœํ•˜๊ธฐ ์œ„ํ•ด ๊ฐœ๋ฐœ๋œ ์•Œ๊ณ ๋ฆฌ ์ฆ˜์œผ๋กœ ์ˆ˜ํ–‰๋œ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์•„์ŠคํŒ”ํŠธ ๋„๋กœํฌ์žฅ ํ‘œ๋ฉด ์˜์ƒ์€ ๋…ธ ์ถœ๋œ ๊ณจ์žฌ, ํ•ด์ƒ๋„, ๋‹จ์ฐจ, ๊ทธ๋ฆผ์ž ๋“ฑ์˜ ์—ฌ๋Ÿฌ ์š”์†Œ๋“ค์— ์˜ํ•ด ๊ฐ ๊ฒฐํ•จ์˜ ํŠน์ง•์„ ๋Œ€ํ‘œํ•˜๋Š” ์ •๋ณด๋ฅผ ๋„์ถœํ•˜๋Š”๋ฐ ์–ด๋ ค์›€์ด ๋”ฐ๋ฅธ ๋‹ค. ๊ฒฐํ•จ์˜ ํŠน์ง•์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ์–ป๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ์˜์ƒ ์ „์ฒ˜๋ฆฌ ์ž‘์—…์ด ์ˆ˜ํ–‰๋˜์–ด์ง€๋‚˜, ์ „์ฒ˜๋ฆฌ ๊ณผ์ •์—์„œ ์˜ค๋Š” ์ •๋ณด ์†์‹ค ๋ฌธ์ œ ์— ์˜ํ•ด ์ž๋™ํ™” ํ‘œ๋ฉด๊ฒฐํ•จ ๋ถ„์„ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ์‹ค๋ฌด์— ์ ์šฉ๋œ ์‚ฌ ๋ก€๋Š” ๋งŽ์ง€ ์•Š๋‹ค. ๊ธฐ์กด์˜ ์˜์ƒ์„ ์‚ฌ์šฉํ•œ ์ž๋™ํ™” ๋„๋กœํฌ์žฅ ํ‘œ๋ฉด ๊ฒฐํ•จ ๋ถ„์„ ๋ฐฉ๋ฒ•์˜ ํ•œ๊ณ„๋ฅผ ๊ทน๋ณตํ•˜๊ธฐ ์œ„ํ•ด ์ตœ๊ทผ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง (CNN; Convolutional Neural Network)์„ ํ™œ์šฉํ•œ ์—ฐ๊ตฌ๋“ค์ด ์ง„ ํ–‰๋˜๊ณ  ์žˆ๋‹ค.

CNN์€ ์˜์ƒ์ฒ˜๋ฆฌ ๋ถ„์•ผ์—์„œ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋‚˜ํƒ€๋‚ด๋Š” ์ธ๊ณต ์‹ ๊ฒฝ๋ง์œผ๋กœ ๋Œ€์šฉ๋Ÿ‰์˜ ์˜์ƒ ๋ฐ์ดํ„ฐ๋กœ๋ถ€ํ„ฐ ๋ถ„๋ฅ˜ํ•˜๊ณ ์ž ํ•˜๋Š” ๋Œ€ ์ƒ์— ๋Œ€ํ•œ ํŠน์ง•์„ ์ž๋™์œผ๋กœ ์ถ”์ถœํ•˜๊ณ , ๊ทธ์— ๋งž๊ฒŒ ์˜์ƒ ๋ฐ์ดํ„ฐ ๋ฅผ ๋ถ„๋ฅ˜ํ•˜๋Š” ๊ธฐ์ˆ ์ด๋‹ค. ํŠน์ • CNN ๋ชจ๋ธ๋“ค์˜ ๊ฒฝ์šฐ ImageNet์— ์„œ ์ œ๊ณตํ•˜๋Š” ์ด๋ฏธ์ง€์— ๋Œ€ํ•ด 97% ์ด์ƒ์˜ ๋ถ„๋ฅ˜ ์ •ํ™•๋„๋ฅผ ๋ณด์—ฌ ์ค€๋‹ค(Hu et al, 2018). ๊ทธ๋Ÿฌ๋‚˜ ์˜์ƒ์ฒ˜๋ฆฌ ๋ถ„์•ผ์˜ CNN ๋ชจ๋ธ ์ด์šฉ ์— ๋”ฐ๋ฅธ ๋†’์€ ๋ถ„๋ฅ˜ ์ •ํ™•๋„ ๊ตฌํ˜„๊ณผ๋Š” ๋‹ค๋ฅด๊ฒŒ ๋„๋กœํฌ์žฅ ํ‘œ๋ฉด ์ด ๋ฏธ์ง€ ๋ถ„๋ฅ˜ ์—ฐ๊ตฌ์—์„œ๋Š” 85~90% ์ •๋„๋กœ ๋‚ฎ์€ ์ •ํ™•๋„๋ฅผ ๋ณด๊ณ ํ•˜ ๊ณ  ์žˆ๋‹ค(Zhang et al., 2016; Feng et al., 2017; Eisenbach et al., 2017; Pauly et al., 2017; Gopalakrishnan, 2018). ๋Œ€๋ถ€๋ถ„์˜ ์—ฐ ๊ตฌ์—์„œ ์‹ญ์—ฌ๊ฐœ ๋ฏธ๋งŒ์˜ ์‹ ๊ฒฝ๋ง ์ธต์„ ๊ฐ–๋Š” CNN ๋ชจ๋ธ์„ ํ™œ์šฉํ•˜ ์˜€์œผ๋ฉฐ, ์ˆ˜์ฒœ~์ˆ˜์‹ญ๋งŒ๊ฐœ ์ •๋„์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šต์— ์‚ฌ์šฉํ•˜์˜€๋‹ค. ๊ธฐ์กด์—ฐ๊ตฌ์˜ CNN ๋ชจ๋ธ๋“ค์€ ๋ณต์žก๋„์— ๋น„ํ•ด ๋ชจ๋ธ ํ•™์Šต ์‹œ ์ƒ๋Œ€ ์ ์œผ๋กœ ์ ์€ ์–‘์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์˜€๊ธฐ ๋•Œ๋ฌธ์— ์ถฉ๋ถ„ํ•œ ์„ฑ๋Šฅ ์„ ๋‚˜ํƒ€๋‚ด์ง€ ๋ชปํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ธ๋‹ค. ์ฆ‰, ๋ณต์žก๋„๊ฐ€ ๋†’์€ CNN ๋ชจ๋ธ์„ ๊ตฌ์ถ•ํ•  ๊ฒฝ์šฐ ๋งŽ์€ ์–‘์˜ ๋ฐ์ดํ„ฐ๊ฐ€ ํ•„์š”ํ•˜๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์•„ ์ŠคํŒ”ํŠธ ์ฝ˜ํฌ๋ฆฌํŠธ ํ‘œ๋ฉด๊ท ์—ด ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ๋Š” ํ•œ์ •์ ์ด๊ธฐ ๋•Œ๋ฌธ ์— ์ ์ ˆํ•œ ๋ณต์žก๋„๋ฅผ ๊ฐ–๋Š” CNN ๋ชจํ˜•์„ ์‚ฌ์šฉํ•ด์•ผ ํ•œ๋‹ค. ๋”ฐ๋ผ์„œ ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ํ™•๋ณด๋œ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ์–‘์— ์ ์ •ํ•œ ๋ณต์žก๋„๋ฅผ ๊ฐ€์ง€๋Š” CNN ๋ชจํ˜•์„ ๊ตฌ์ถ•ํ•˜์˜€์œผ๋ฉฐ, ํ•™์Šต ๋ฐ์ดํ„ฐ์–‘์— ๋”ฐ๋ฅธ ์„ฑ ๋Šฅํ–ฅ์ƒ ์ •๋„๋ฅผ ํ‰๊ฐ€ํ•˜๊ณ ์ž ํ•œ๋‹ค.

2. ์ด๋ก ์  ๋ฐฐ๊ฒฝ

ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง์€ ์ž„์˜์˜ ํ•„ํ„ฐ๋ฅผ ํ†ตํ•ด ์˜์ƒ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ํŠน์ง•์„ ์ถ”์ถœํ•˜์—ฌ ๋™์ผํ•œ ๋ ˆ์ด๋ธ”์„ ์ง€๋‹Œ ๋ฐ์ดํ„ฐ๋“ค์˜ ๊ณตํ†ต์ ์ธ ํŠน์ง•์— ๋Œ€ํ•ด ํ•™์Šต์„ ์ง„ํ–‰ํ•˜๋Š” ์ธ๊ณต์‹ ๊ฒฝ๋ง์ด๋‹ค.

ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง์˜ ๊ตฌ์กฐ๋Š” Fig. 1๊ณผ ๊ฐ™์ด ํฌ๊ฒŒ Convolution Layer์™€ Pooling ๊ณผ์ •์ด ํฌํ•จ๋œ ํŠน์ง• ํ•™์Šต ์‹ ๊ฒฝ๋ง ๋ถ€๋ถ„(Featu- -re Learning)๊ณผ ๋ ˆ์ด๋ธ”์— ๋”ฐ๋ฅธ ๋ถ„๋ฅ˜๋ฅผ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ถ„๋ฅ˜ ์‹ ๊ฒฝ๋ง (Classification) ๋ถ€๋ถ„์œผ๋กœ ๋‚˜๋‰œ๋‹ค(Wu 2017; Choi 2019).

Fig. 1

Schematic of CNN(Convolutional Neural Network) architecture

JKSMI-23-6-38_F1.jpg

ํŠน์ง• ํ•™์Šต ์‹ ๊ฒฝ๋ง์€ ํ•„ํ„ฐ๋ฅผ ํ†ตํ•ด ํŠน์ง• ๋ฒกํ„ฐ๋ฅผ ์ถ”์ถœํ•˜๋Š” ๊ตฌ ๊ฐ„์ด๋‹ค. ํ•„ํ„ฐ๋Š” ์ดˆ๊ธฐ ์ž„์˜์˜ ๊ฐ’์œผ๋กœ ์ฃผ์–ด์ง€๋ฉฐ ํ•„ํ„ฐ์˜ ํฌ๊ธฐ๋กœ ์ „์ฒด์˜ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ๋ฅผ ์Šค์บ”ํ•˜์—ฌ ํŠน์ง• ๋ฒกํ„ฐ๋ฅผ ์ถ”์ถœํ•ด๋‚ธ๋‹ค. Fig. 2๋Š” 3x3 ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด 1x1 Stride ํฌ๊ธฐ์™€ 2x2 ํ•„ ํ„ฐ๋ฅผ ์ ์šฉํ•œ Convolution์˜ ์˜ˆ์‹œ์ด๋‹ค. ์—ฌ๊ธฐ์„œ, Stride๋Š” ํ•„ํ„ฐ ์— ์˜ํ•œ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ ์Šค์บ”์ด ์ด๋ฃจ์–ด์งˆ ๋•Œ์˜ ๊ฐ„๊ฒฉ์˜ ํฌ๊ธฐ๋ฅผ ์˜๋ฏธํ•œ๋‹ค. Fig. 2์™€ ๊ฐ™์ด convolution์— ์˜ํ•ด ์‚ฐ์ถœ๋œ ํŠน์ง• ๋ฒกํ„ฐ ์˜ ํฌ๊ธฐ๋Š” ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ์˜ ํฌ๊ธฐ๋ณด๋‹ค ์ž‘์•„์ง€๋Š” ๊ฒƒ์„ ๋ณผ ์ˆ˜ ์žˆ ๋‹ค. convolution์„ ์ง„ํ–‰ํ•  ๋•Œ๋งˆ๋‹ค ํŠน์ง• ๋ฒกํ„ฐ์˜ ํฌ๊ธฐ๊ฐ€ ์ž‘์•„์ง€ ๋Š” ํ˜„์ƒ ๋•Œ๋ฌธ์— layer ๊นŠ์ด์˜ ํ•œ๊ณ„๊ฐ€ ๋‚˜ํƒ€๋‚˜๊ฒŒ ๋œ๋‹ค. ์ด๋Ÿฌํ•œ ํ˜„ ์ƒ์„ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•œ ๋ฐฉ๋ฒ•์œผ๋กœ Padding์ด ์žˆ๋‹ค. Padding์€ ์ด ๋ฏธ์ง€ ๋ฒกํ„ฐ์˜ ๋ ํ–‰๊ณผ ์—ด์— ์˜๋ฒกํ„ฐ๋ฅผ ์ถ”๊ฐ€ํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ convolution์˜ ์—ฐ์‚ฐ๊ณผ์ • ํ›„ ์‚ฐ์ถœ๋˜๋Š” ํŠน์ง• ๋ฒกํ„ฐ์™€ ์ž…๋ ฅ ์ด๋ฏธ ์ง€ ๋ฒกํ„ฐ์˜ ํฌ๊ธฐ๋ฅผ ๋™์ผํ•˜๊ฒŒ ๋งŒ๋“ค์–ด์ฃผ๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•œ๋‹ค. convolution ์—ฐ์‚ฐ๊ณผ์ •์„ ๊ฑฐ์ณ ์‚ฐ์ถœ๋œ ํŠน์ง• ๋ฒกํ„ฐ๋Š” ํ™œ์„ฑํ™” ํ•จ์ˆ˜์ธ ReLU ํ•จ์ˆ˜์˜ ์ž…๋ ฅ๊ฐ’์œผ๋กœ ์ ์šฉ๋œ๋‹ค(Radford 2015).

Fig. 2

Example of convolution

JKSMI-23-6-38_F2.jpg

Convolution Layer๋ฅผ ๊ฑฐ์ณ ์‚ฐ์ถœ๋œ ํŠน์ง• ๋ฒกํ„ฐ๋Š” Pooling ๊ณผ ์ •์„ ๊ฑฐ์นœ๋‹ค. Pooling์˜ ์ข…๋ฅ˜๋Š” Max Pooling, Average Pooling ๋“ฑ์œผ๋กœ ๋‹ค์–‘ํ•˜๊ฒŒ ์กด์žฌํ•œ๋‹ค(Goodfellow 2016).

๋Œ€๋ถ€๋ถ„์˜ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง์—์„œ๋Š” Max Pooling์„ ์‚ฌ์šฉํ•˜๋ฉฐ, Fig. 3๊ณผ ๊ฐ™์ด ์ปค๋„(kernel) ํฌ๊ธฐ ๋‚ด์— ์กด์žฌํ•˜๋Š” ๋ฒกํ„ฐ์˜ ๊ฐ’ ์ค‘ ๊ฐ€์žฅ ํฐ ๊ฐ’์„ ์ถ”์ถœํ•ด๋‚ด๋Š” Pooling ๋ฐฉ์‹์„ ์˜๋ฏธํ•œ๋‹ค.

Fig. 3

Example of Max Pooling

JKSMI-23-6-38_F3.jpg

ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์˜ layer ์ˆ˜๋งŒํผ Convolutional Layer์™€ Pooling ๊ณผ์ •์„ ๋ฐ˜๋ณตํ•˜์—ฌ ํŠน์ง• ๋ฒกํ„ฐ๋ฅผ ๋„์ถœํ•œ๋‹ค. ๋„์ถœ๋œ ํŠน์ง• ๋ฒกํ„ฐ๋ฅผ FCNN(Fully-Connected Neural Network)์˜ ์ž…๋ ฅ๊ฐ’์œผ ๋กœ ํ™œ์šฉํ•˜์—ฌ FCNN์„ ํ†ตํ•ด ์ดˆ๊ธฐ ์˜ˆ์ธก๋œ ์ถœ๋ ฅ๊ฐ’์„ ์‚ฐ์ถœํ•œ๋‹ค. ์ตœ์ข…์ ์œผ๋กœ, ์˜ˆ์ธก๋œ ์ถœ๋ ฅ๊ฐ’๊ณผ ์‹ค์ œ ๋ ˆ์ด๋ธ” ๊ฐ’์˜ ์ฐจ์ด๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ์˜ ์ •ํ™•๋„๋ฅผ ์ธก์ •ํ•˜๋Š” ๋น„์šฉ ํ•จ์ˆ˜(Loss Function)๋ฅผ ์‚ฌ์šฉ ํ•˜์—ฌ loss๊ฐ’์„ ์‚ฐ์ถœํ•œ๋‹ค(Bengio et al. 2015). ํ•™์Šต์˜ ์˜๋ฏธ๋Š” ์‚ฐ ์ถœ๋˜๋Š” loss์˜ ๊ฐ’์„ ์ตœ์†Œํ™”์‹œํ‚ค๋Š” ํ•„ํ„ฐ ๋ฒกํ„ฐ๋ฅผ ๋„์ถœํ•˜์—ฌ ์‹ค์ œ ๋ ˆ์ด๋ธ” ๊ฐ’๊ณผ ์˜ˆ์ธก๋œ ์ถœ๋ ฅ๊ฐ’์˜ ์ฐจ์ด๋ฅผ ์ค„์ด๋Š” ๊ฒƒ์ด๋‹ค. loss ๊ฐ’ ์„ ์ค„์ด๋Š” ํ•™์Šต ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ ๊ฒฝ์‚ฌํ•˜๊ฐ•๋ฒ•(Gradient Descent) ์ด ์žˆ์œผ๋ฉฐ, ๊ฒฝ์‚ฌํ•˜๊ฐ•๋ฒ•์„ ํ†ตํ•ด loss๊ฐ’์„ ๊ฐ์†Œ์‹œํ‚ค๋Š” ํ•„ํ„ฐ ๋ฒกํ„ฐ ๋กœ ๊ฐฑ์‹ ํ•œ๋‹ค.

3. ๋ณธ ๋ก 

3.1 ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ

์•ž์„  ์ด๋ก ์  ๋ฐฐ๊ฒฝ์„ ๋ฐ”ํƒ•์œผ๋กœ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์„ ๊ตฌ์ถ•ํ•˜ ์˜€์œผ๋ฉฐ, ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ์–‘์„ ์ผ์ • ํฌ๊ธฐ(batch size)๋กœ ๋‚˜๋ˆ„์–ด ํ•™์Šต ํ•˜๋Š” Mini-batch(Keskar et al., 2016)์™€ ๋ฐฐ์น˜ ์ •๊ทœํ™”(Batch Normalization)(Loffe and Szegedy, 2015)๋ฅผ ์ ์šฉํ•˜์—ฌ ํ•™์Šต ์„ฑ๋Šฅ ๋ฐ ์†๋„๋ฅผ ํ–ฅ์ƒ์‹œ์ผฐ๋‹ค. ๋ณธ ์—ฐ๊ตฌ์˜ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์— ์  ์šฉํ•œ Filter size, stride size, batch size ๋“ฑ์„ Table 1์— ์ •๋ฆฌํ•˜์˜€์œผ ๋ฉฐ, ํ•™์Šต ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ๊ฒฝ์‚ฌํ•˜๊ฐ•๋ฒ•์„ ์‚ฌ์šฉํ•˜์˜€๋‹ค. ๋ชจ๋ธ์˜ ๊ตฌ์กฐ๋Š” Fig. 4์™€ ๊ฐ™์ด 5๊ฐœ ์ธต์˜ Convolution Layer๋กœ ๊ตฌ์„ฑํ•˜์˜€๋‹ค.

Table 1

Parameter of CNN used in this study

JKSMI-23-6-38_T1.jpg
Fig. 4

The Architecture of CNN used in this study

JKSMI-23-6-38_F4.jpg

3.2 ํ•™์Šต ๋ฐ ์‹œํ—˜ ๋ฐ์ดํ„ฐ

๋ณธ ์—ฐ๊ตฌ์—์„œ ํ•™์Šต ๋ฐ ์‹œํ—˜ ๋ฐ์ดํ„ฐ๋กœ ํ™œ์šฉ๋œ ์•„์ŠคํŒ”ํŠธ ์ฝ˜ํฌ ๋ฆฌํŠธ ๋„๋กœํฌ์žฅ ์˜์ƒ ๋ฐ์ดํ„ฐ๋Š” ๊ตญ๋‚ด 5๊ฐœ ์ง€์—ญ์˜ ๊ตญ๋„์—์„œ KRISS ์žฅ๋น„๋ฅผ ํ†ตํ•ด ์กฐ์‚ฌ๋œ ๋ฐ์ดํ„ฐ์ด๋‹ค. ํ•™์Šต ๋ฐ ์‹œํ—˜ ๋ฐ์ดํ„ฐ ๋Š” ๊ฐ™์€ ๋ชจ์ง‘๋‹จ์—์„œ 7:3 ํ˜น์€ 8:2์˜ ๋น„์œจ์— ๋งž๊ฒŒ ๋‚˜๋ˆ„์–ด ๊ตฌ์ถ• ํ•˜๋Š” ๊ฒƒ์ด ์ผ๋ฐ˜์ ์ด๋‚˜, ๋ณธ ์—ฐ๊ตฌ์˜ ํ•™์Šต ๋ฐ ์‹œํ—˜ ๋ฐ์ดํ„ฐ๋Š” ๊ฐ™์€ ๋ชจ์ง‘๋‹จ์—์„œ ๊ตฌ์ถ•ํ•˜์ง€ ์•Š์•˜๋‹ค. ๋ณธ ์—ฐ๊ตฌ์˜ ํ•™์Šต ๋ฐ์ดํ„ฐ๋Š” 4๊ฐœ ์ง€์—ญ์˜ ๊ตญ๋„ ์˜์ƒ ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•˜์˜€์œผ๋ฉฐ, ์‹œํ—˜ ๋ฐ์ดํ„ฐ์˜ ๊ฒฝ ์šฐ ํ•™์Šต ๋ฐ์ดํ„ฐ์— ํฌํ•จ๋˜์ง€ ์•Š์€ 1๊ฐœ ์ง€์—ญ์˜ ๊ตญ๋„ ์˜์ƒ ๋ฐ์ด ํ„ฐ๋ฅผ ํ™œ์šฉํ•˜์˜€๋‹ค. KRISS ์žฅ๋น„๋ฅผ ํ†ตํ•ด ์กฐ์‚ฌ๋œ ๋…ธ๋ฉด ์˜์ƒ ๋ฐ์ด ํ„ฐ์˜ ํฌ๊ธฐ๋Š” ํญ 3.6m, ๊ธธ์ด 10m์— ํ•ด๋‹นํ•˜๋ฉฐ, ํ•™์Šต ๋ฐ ์‹œํ—˜ ๋ฐ ์ดํ„ฐ๋Š” ์กฐ์‚ฌ๋œ ๋…ธ๋ฉด ์˜์ƒ ๋ฐ์ดํ„ฐ๋ฅผ ํญ 300mm, ๊ธธ์ด 300mm์˜ ํฌ๊ธฐ๋กœ ๋ถ„ํ• ํ•œ 300x300 ํ”ฝ์…€์˜ ์ด๋ฏธ์ง€์ด๋‹ค. ํ•™์Šต ๋ฐ ์‹œํ—˜ ๋ฐ ์ดํ„ฐ์˜ ๊ตฌ์„ฑ์€ ํ‘œ๋ฉด๊ท ์—ด ๋ฐœ์ƒ ์—ฌ๋ถ€์— ๋”ฐ๋ผ ๊ท ์—ด๊ณผ ๋น„๊ท ์—ด ๋ฐ ์ดํ„ฐ๋กœ ๋‚˜๋ˆ  ๊ตฌ์ถ•ํ•˜์˜€๋‹ค. Fig. 5(a)๋Š” ํ•™์Šต ๋ฐ ์‹œํ—˜ ๋ฐ์ดํ„ฐ๋กœ ํ™œ์šฉ๋œ ๊ท ์—ด ๋ฐ์ดํ„ฐ, Fig. 5(b)๋Š” ๋น„๊ท ์—ด ๋ฐ์ดํ„ฐ์˜ ์ผ๋ถ€๋ฅผ ๋ณด์—ฌ ์ค€๋‹ค. CNN ๋ชจ๋ธ ํ•™์Šต ์‹œ Fig. 5 ์™€ ๊ฐ™์€ ์ด๋ฏธ์ง€ ์ •๋ณด๊ฐ€ ์ž…๋ ฅ๊ฐ’ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋ฉฐ, ๊ท ์—ด ๋ฐ ๋น„๊ท ์—ด์„ ๋Œ€ํ‘œํ•˜๋Š” ๊ฐ’์ด ์ถœ๋ ฅ๊ฐ’์œผ๋กœ ์‚ฌ์šฉ๋œ๋‹ค.

Fig. 5

Sample of Training and Test Data

JKSMI-23-6-38_F5.jpg

ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์˜ ๋ถ„๋ฅ˜ ๊ฒ€์ถœ ์„ฑ๋Šฅ ์ฆ์ง„์— ์žˆ์–ด ๋ฐ์ดํ„ฐ ์˜ ์–‘์€ ๋งค์šฐ ์ค‘์š”ํ•œ ์š”์†Œ์ด๋‹ค. ๋”ฐ๋ผ์„œ ๋ฐ์ดํ„ฐ์˜ ์–‘์„ ์ฆ์ง„์‹œ ํ‚ค๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•์œผ๋กœ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ๋ฅผ ๊ฐ€๊ณตํ•œ๋‹ค. ์ด๋ฏธ ์ง€ ๋ณ€ํ™˜์˜ ๋ฐฉ๋ฒ•์—๋Š” ์ž˜๋ผ๋‚ด๊ธฐ, ํšŒ์ „, ๋’ค์ง‘๊ธฐ, ๋ฐ€๊ธฐ, ํฌ๊ธฐ๋ณ€ํ™˜ ๋“ฑ์ด ์žˆ๋‹ค. ๋ณธ ์—ฐ๊ตฌ์—์„œ ์›๋ณธ ๋ฐ์ดํ„ฐ๋ฅผ 90๋„, 180๋„, 270๋„ ํšŒ ์ „์‹œ์ผœ ๋ฐ์ดํ„ฐ์˜ ์–‘์„ ์ฆ์ง„์‹œ์ผฐ๋‹ค.

์ผ๋ฐ˜์ ์œผ๋กœ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ๋Š” ์—ฌ๋Ÿฌ ๋ณ€์ˆ˜์— ์˜ํ•ด ํ•ด์ƒ๋„, ๋ฐ ๊ธฐ ๋“ฑ์ด ์ƒ์ดํ•˜๋‹ค. ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ์˜ ์ผ๊ด€์„ฑ์„ ์œ„ํ•ด ์ด๋ฏธ์ง€ ์ „ ์ฒ˜๋ฆฌ ์ž‘์—…์ด ์š”๊ตฌ๋œ๋‹ค. ๋ณธ ์—ฐ๊ตฌ์—์„œ ์‚ฌ์šฉํ•œ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ์˜ ๊ฐ ํ”ฝ์…€ ๊ฐ’์€ 0~255 ๋ฒ”์œ„ ๋‚ด์˜ ๊ฐ’์„ ๊ฐ–๋Š”๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ•™์Šตํ•˜๊ณ  ์ž ํ•˜๋Š” ๋ชจ๋“  ์ด๋ฏธ์ง€ ๋ฒกํ„ฐ์˜ ์ตœ์†Œ, ์ตœ๋Œ“๊ฐ’์ด 0๊ณผ 255์— ํ•ด๋‹นํ•˜ ์ง€ ์•Š์œผ๋ฉฐ, ๊ฐ’์˜ ๋ฒ”์œ„๊ฐ€ ๋‹ค๋ฅด๊ณ  ๋‹ค์–‘ํ•˜๊ฒŒ ๋ถ„ํฌ๋˜์–ด์žˆ๋Š” ์ด๋ฏธ ์ง€ ๋ฐ์ดํ„ฐ๋“ค์— ๋Œ€ํ•œ ํ•™์Šต์„ ์ง„ํ–‰ํ•˜๋Š”๋ฐ ์–ด๋ ค์›€์ด ๋”ฐ๋ฅธ๋‹ค. ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ์‹ (1)๊ณผ ๊ฐ™์€ ์Šค์ผ€์ผ ์ •๊ทœํ™” ์‹์„ ์ ์šฉํ•˜์—ฌ ๋ชจ๋“  ์ด๋ฏธ์ง€ ๋ฒกํ„ฐ์˜ ์ตœ์†Œ, ์ตœ๋Œ“๊ฐ’์„ [0,1]์ด ๋˜๋„๋ก ํ•˜์˜€๋‹ค.

(1)
X = ( x i โˆ’ x min ) * ( max - min ) x max โˆ’ x min + min

์—ฌ๊ธฐ์„œ, X๋Š” ์Šค์ผ€์ผ ์ •๊ทœํ™”๊ฐ€ ์ ์šฉ๋œ ์ด๋ฏธ์ง€ ๋ฒกํ„ฐ, xi๋Š” ๊ธฐ ์กด ์ด๋ฏธ์ง€ ๋ฒกํ„ฐ์˜ ์ž„์˜๊ฐ’, xmin๋Š” ๊ธฐ์กด ์ด๋ฏธ์ง€ ๋ฒกํ„ฐ์˜ ์ตœ์†Ÿ๊ฐ’, xmax๋Š” ๊ธฐ์กด ์ด๋ฏธ์ง€ ๋ฒกํ„ฐ์˜ ์ตœ๋Œ“๊ฐ’, max๋Š” ๋ณ€ํ™˜ํ•˜๊ณ ์ž ํ•˜๋Š” ๋ฒ”์œ„์˜ ์ตœ๋Œ“๊ฐ’, min์€ ๋ณ€ํ™˜ํ•˜๊ณ ์ž ํ•˜๋Š” ๋ฒ”์œ„์˜ ์ตœ์†Ÿ๊ฐ’์ด๋‹ค.

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ์–‘์— ๋”ฐ๋ฅธ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง ๋ชจ ๋ธ์˜ ์„ฑ๋Šฅ๋ณ€ํ™”๋ฅผ ํ™•์ธํ•˜๊ธฐ ์œ„ํ•ด 6๊ฐœ์˜ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ ์„ฑํ•˜์˜€๋‹ค. ํ•™์Šต ๋ฐ์ดํ„ฐ ๋ฐ ์‹œํ—˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ตฌ์„ฑ์€ Table 2 ์™€ ๊ฐ™๋‹ค.

Table 2

Composition of training and test data

JKSMI-23-6-38_T2.jpg

Table 2์˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋ผ๋ฒจ์˜ T๋Š” ์›๋ณธ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ๋ฅผ ์˜๋ฏธํ•˜๊ณ , TA๋Š” ์›๋ณธ ์ด๋ฏธ์ง€๋ฅผ ๋ณ€ํ™˜ํ•˜์—ฌ ๊ตฌ์ถ•ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ ์ด๋‹ค. ๋˜ํ•œ, ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋ผ๋ฒจ์˜ ์ˆซ์ž๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ํฌํ•จ๋œ ๊ท ์—ด๊ณผ ๋น„๊ท ์—ด ์ด๋ฏธ์ง€ ์ˆ˜์˜ ํ•ฉ์„ ์˜๋ฏธํ•œ๋‹ค. Aug.๋Š” ์ด๋ฏธ์ง€ ๋ณ€ ํ™˜ ๋ฐฉ์‹์„ ํ†ตํ•ด ์ƒ์„ฑํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ์˜๋ฏธํ•˜๋ฉฐ, None_Aug.๋Š” ์› ๋ณธ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ๋ฅผ ์˜๋ฏธํ•œ๋‹ค. ๋ณธ ์—ฐ๊ตฌ์˜ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ ์€ Table 2์˜ ๊ฐ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ํ•™์Šต์„ ์ง„ํ–‰ํ•˜์˜€๋‹ค.

3.3 ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์˜ ์„ฑ๋Šฅํ‰๊ฐ€

ํ•™์Šต๋œ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์˜ ์„ฑ๋Šฅํ‰๊ฐ€๋Š” ์‹œํ—˜ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๊ท ์—ด ๊ฒ€์ถœ ์ •ํ™•๋„(Accuracy), ์ •๋ฐ€๋„(Precision), ์žฌํ˜„์œจ (Recall), ๋ฏธ๊ฒ€์ถœ์œจ(Missing Rate), ๊ณผ๊ฒ€์ถœ์œจ(Over Rate)์„ ์‚ฐ ์ถœ์„ ํ†ตํ•ด ์ง„ํ–‰ํ•˜์˜€๋‹ค.

๊ฒ€์ถœ ์ •ํ™•๋„๋Š” ์ „์ฒด ์ด๋ฏธ์ง€ ์ค‘์—์„œ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์ด ์ •ํ™•ํ•˜๊ฒŒ ๊ท ์—ด๊ณผ ๋น„๊ท ์—ด์„ ํŒ๋‹จํ•œ ๊ฐœ์ˆ˜์˜ ๋น„์œจ์„ ์˜๋ฏธํ•˜๋ฉฐ, ์‹ (2)๋ฅผ ํ†ตํ•ด ์‚ฐ์ถœ๋œ๋‹ค.

(2)
A c c = T P + T N T P + T N + F P + F N

์—ฌ๊ธฐ์„œ, Acc๋Š” ๊ฒฐํ•จ ๊ฒ€์ถœ ์ •ํ™•๋„,TP(True Positive)๋Š” ๊ท ์—ด ์ด๋ฏธ์ง€๋ฅผ ๊ท ์—ด๋กœ ํŒ๋‹จํ•œ ๊ฐœ์ˆ˜,TN(True Negative)์€ ๋น„๊ท ์—ด ์ด๋ฏธ์ง€๋ฅผ ๋น„๊ท ์—ด๋กœ ํŒ๋‹จํ•œ ๊ฐœ์ˆ˜, FP(False Positive)๋Š” ๋น„๊ท ์—ด ์ด๋ฏธ์ง€๋ฅผ ๊ท ์—ด๋กœ ํŒ๋‹จํ•œ ๊ฐœ์ˆ˜, FN(False Negative)์€ ๊ท ์—ด ์ด ๋ฏธ์ง€๋ฅผ ๋น„๊ท ์—ด๋กœ ํŒ๋‹จํ•œ ๊ฐœ์ˆ˜์ด๋‹ค.

๊ฒ€์ถœ ์ •ํ™•๋„๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋ถ„๋ฅ˜ํ•˜๊ณ ์žํ•˜๋Š” ๋ฐ์ดํ„ฐ๊ฐ€ ๊ท ๋“ฑ ํ•œ ๋ถ„ํฌ๋กœ ๊ตฌ์„ฑ๋˜์–ด์žˆ์„ ๋•Œ ๊ธฐ์ดˆ์ ์œผ๋กœ ์‚ฌ์šฉ๋œ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ถˆ ๊ท ๋“ฑํ•œ ๋ถ„ํฌ๋ฅผ ๊ฐ–๋Š” ๋ฐ์ดํ„ฐ ๊ตฌ์„ฑ์˜ ๊ฒฝ์šฐ ๋‹จ์ˆœํžˆ ๊ฒ€์ถœ ์ •ํ™•๋„ ๋งŒ์œผ๋กœ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•  ์ˆ˜ ์—†๋‹ค. ๋”ฐ๋ผ์„œ ์ •ํ™•ํ•œ ๋ถ„๋ฅ˜ ๊ฒ€์ถœ๊ธฐ์˜ ์„ฑ๋Šฅ์— ๋Œ€ํ•œ ์ง€ํ‘œ๋Š” ์ •ํ™•๋„๋ฅผ ๋”ํ•ด ์ •๋ฐ€๋„์™€ ์žฌํ˜„์œจ์ด ํ™œ์šฉ ๋œ๋‹ค(Powers 2011). ์ผ๋ฐ˜์ ์œผ๋กœ ์•„์ŠคํŒ”ํŠธ ์ฝ˜ํฌ๋ฆฌํŠธ ๋„๋กœํฌ์žฅ ํ‘œ๋ฉด๊ฒฐํ•จ์€ ๋ถˆ๊ท ๋“ฑํ•˜๊ฒŒ ๊ท ์—ด์ด ๋ถ„ํฌ๋˜์–ด์žˆ์œผ๋ฏ€๋กœ ๊ท ์—ด ๊ฒ€์ถœ ์ •ํ™•๋„์™€ ์ •๋ฐ€๋„, ์žฌํ˜„์œจ์˜ ์„ฑ๋Šฅ์„ ํ™•์ธํ•˜์˜€์œผ๋ฉฐ, ์ถ”๊ฐ€์ ์œผ ๋กœ ๋ฏธ๊ฒ€์ถœ์œจ๊ณผ ๊ณผ๊ฒ€์ถœ์œจ์— ๋Œ€ํ•œ ์„ฑ๋Šฅ์„ ํ™•์ธํ•˜์˜€๋‹ค.

์ •๋ฐ€๋„๋Š” ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์ด ๊ท ์—ด๋กœ ํŒ๋‹จํ•œ ์ด๋ฏธ์ง€ ์ค‘ ์—์„œ ์‹ค์ œ ๊ท ์—ด์ด ์กด์žฌํ•˜๋Š” ์ด๋ฏธ์ง€์˜ ๊ฐœ์ˆ˜์— ๋Œ€ํ•œ ๋น„์œจ์„ ์˜ ๋ฏธํ•˜๋ฉฐ, ์‹ (3)์„ ํ†ตํ•ด ์‚ฐ์ถœ๋œ๋‹ค.

(3)
P r e c = T P T P + F P

์—ฌ๊ธฐ์„œ, Prec๋Š” ๊ท ์—ด ๊ฒ€์ถœ ์ •๋ฐ€๋„, TP๋Š” ๊ท ์—ด ์ด๋ฏธ์ง€๋ฅผ ๊ท  ์—ด๋กœ ํŒ๋‹จํ•œ ๊ฐœ์ˆ˜, FP๋Š” ๋น„๊ท ์—ด ์ด๋ฏธ์ง€๋ฅผ ๊ท ์—ด๋กœ ํŒ๋‹จํ•œ ๊ฐœ์ˆ˜ ์ด๋‹ค.

์žฌํ˜„์œจ์ด๋ž€ ์‹ค์ œ ๊ท ์—ด์ด ์กด์žฌํ•˜๋Š” ์ด๋ฏธ์ง€ ์ค‘์—์„œ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์ด ์ •ํ™•ํ•˜๊ฒŒ ๊ท ์—ด๋กœ ํŒ๋‹จํ•œ ๊ฐœ์ˆ˜์˜ ๋น„์œจ์„ ์˜๋ฏธ ํ•˜๋ฉฐ, ์‹ (4)๋ฅผ ํ†ตํ•ด ์‚ฐ์ถœ๋œ๋‹ค.

(4)
R e c = T P T P + F N

์—ฌ๊ธฐ์„œ, Recf๋Š” ๊ท ์—ด ๊ฒ€์ถœ ์žฌํ˜„์œจ, TP๋Š” ๊ท ์—ด ์ด๋ฏธ์ง€๋ฅผ ๊ท  ์—ด๋กœ ํŒ๋‹จํ•œ ๊ฐœ์ˆ˜, FN์€ ๊ท ์—ด ์ด๋ฏธ์ง€๋ฅผ ๋น„๊ท ์—ด๋กœ ํŒ๋‹จํ•œ ๊ฐœ ์ˆ˜์ด๋‹ค.

๋ฏธ๊ฒ€์ถœ์œจ์€ ์‹ค์ œ ๊ท ์—ด์ด ์กด์žฌํ•˜๋Š” ์ด๋ฏธ์ง€ ์ค‘์—์„œ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์ด ๋น„๊ท ์—ด๋กœ ํŒ๋‹จํ•œ ๊ฐœ์ˆ˜์˜ ๋น„์œจ์„ ์˜๋ฏธํ•˜๋ฉฐ, ์‹ (5)๋ฅผ ํ†ตํ•ด ์‚ฐ์ถœ๋œ๋‹ค.

(5)
M = F N A P

์—ฌ๊ธฐ์„œ, M์€ ๊ท ์—ด ๋ฏธ๊ฒ€์ถœ์œจ, FN์€ ๊ท ์—ด ์ด๋ฏธ์ง€๋ฅผ ๋น„๊ท ์—ด ๋กœ ํŒ๋‹จํ•œ ๊ฐœ์ˆ˜, AP๋Š” ์‹ค์ œ ๊ท ์—ด ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ์˜ ๊ฐœ์ˆ˜์ด๋‹ค.

๊ณผ๊ฒ€์ถœ์œจ์€ ๋น„๊ท ์—ด ์ด๋ฏธ์ง€ ์ค‘์—์„œ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์ด ๊ท  ์—ด๋กœ ํŒ๋‹จํ•œ ๊ฐœ์ˆ˜์˜ ๋น„์œจ์„ ์˜๋ฏธํ•˜๋ฉฐ, ์‹ (6)์„ ํ†ตํ•ด ์‚ฐ์ถœ๋œ๋‹ค.

(6)
O = F P A N

์—ฌ๊ธฐ์„œ, O๋Š” ๊ท ์—ด ๊ณผ๊ฒ€์ถœ์œจ, FP๋Š” ๋น„๊ท ์—ด ์ด๋ฏธ์ง€๋ฅผ ๊ท ์—ด๋กœ ํŒ๋‹จํ•œ ๊ฐœ์ˆ˜, AN์€ ๋น„๊ท ์—ด ์ด๋ฏธ์ง€์˜ ๊ฐœ์ˆ˜์ด๋‹ค.

4. ์‹คํ—˜๊ฒฐ๊ณผ ๋ฐ ๊ณ ์ฐฐ

ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์„ ์ด์šฉํ•œ ์•„์ŠคํŒ”ํŠธ ์ฝ˜ํฌ๋ฆฌํŠธ ๋„๋กœํฌ์žฅ ํ‘œ๋ฉด๊ท ์—ด ๊ฒ€์ถœ์— ๋Œ€ํ•œ ์‹คํ—˜์€ ๊ฐ๊ธฐ ๋‹ค๋ฅธ ํ•™์Šต ๋ฐ์ดํ„ฐ ๊ตฌ์„ฑ์„ ํ•™์Šตํ•œ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์„ ํ†ตํ•ด ๊ฐ™์€ ์‹œํ—˜ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๊ฒ€์ถœ ์„ฑ๋Šฅํ‰๊ฐ€๋ฅผ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค. ์‹œํ—˜ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์„ฑ๋Šฅ ํ‰๊ฐ€ ๊ฒฐ๊ณผ๋Š” Table 3๊ณผ ๊ฐ™๋‹ค.

Table 3

Classification performance of test data of CNN model with increase of training data size

JKSMI-23-6-38_T3.jpg

Table 3์„ ํ†ตํ•ด ์•Œ ์ˆ˜ ์žˆ๋“ฏ์ด ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ์–‘์ด ์ฆ๊ฐ€๋จ์— ๋”ฐ ๋ผ ๋ชจ๋“  ์„ฑ๋Šฅ์ง€ํ‘œ๊ฐ€ ํ–ฅ์ƒ๋˜๋Š” ๊ฒƒ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค. ํŠนํžˆ ๋‚ฎ์€ ์ˆ˜์น˜์ผ์ˆ˜๋ก ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์ธ ๊ณผ๊ฒ€์ถœ์œจ ๋ฐ ๋ฏธ๊ฒ€์ถœ์œจ์˜ ๊ฐ์†Œ ํญ์€ ํฌ๊ฒŒ ๋ฐœ์ƒํ•˜๋Š” ๊ฒƒ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด T6000๊ณผ TA96000์„ ํ•™์Šตํ•œ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง์˜ ๊ณผ๊ฒ€์ถœ์œจ๊ณผ ๋ฏธ๊ฒ€์ถœ์œจ์„ ๋น„ ๊ตํ•˜์˜€์„ ๋•Œ ๊ฐ๊ฐ ์•ฝ 2๋ฐฐ ๋ฐ 4๋ฐฐ ์ด์ƒ์˜ ์ฐจ์ด๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ๊ฒƒ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ•™์Šต ๋ฐ์ดํ„ฐ ์ˆ˜์— ๋”ฐ๋ฅธ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง์˜ ์„ฑ๋Šฅํ–ฅ์ƒ์€ Fig. 6๊ณผ Fig. 7์—์„œ ๋ณด๋‹ค ๋ช…ํ™•ํ•˜๊ฒŒ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค.

Fig. 6

Comparison of Classification Detection Accuracy, Precision, Recall for Test Data according to composition of Training Data

JKSMI-23-6-38_F6.jpg
Fig. 7

Comparison of Classification Detection Missing rate, Over rate for Test Data according to composition of Training Data

JKSMI-23-6-38_F7.jpg

์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์˜ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ถ”๊ฐ€์ ์ธ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ์œ„ํ•ด์„œ ํ•™์Šต ๋ฐ์ดํ„ฐ์–‘์˜ ์ฆ์ง„ ํญ์ด ๋” ํฌ๊ฒŒ ์š”๊ตฌ๋œ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด Fig. 6์—์„œ ์ •ํ™•๋„ 95.07%์—์„œ 96.92%๋กœ ์•ฝ 2% ์ฆ ๊ฐ€์‹œํ‚ค๊ธฐ ์œ„ํ•œ ํ•™์Šต ๋ฐ์ดํ„ฐ ์ˆ˜๋Š” 12000์—์„œ 24000์œผ๋กœ 2๋ฐฐ๊ฐ€ ์š”๊ตฌ๋˜๋Š” ๋ฐ˜๋ฉด 96.92%์—์„œ 97.46%๋กœ ์•ฝ 0.5% ์ฆ๊ฐ€์‹œํ‚ค๊ธฐ ์œ„ ํ•œ ํ•™์Šต ๋ฐ์ดํ„ฐ ์ˆ˜๋Š” 24000์—์„œ 96000์œผ๋กœ 4๋ฐฐ๊ฐ€ ์š”๊ตฌ๋œ๋‹ค.

Fig. 8๊ณผ Fig. 9๋Š” ์›๋ณธ ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ์ถ•ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ(T1200 0, T24000)์™€ ์ด๋ฏธ์ง€ ๋ณ€ํ™˜์„ ํ†ตํ•ด ๋ฐ์ดํ„ฐ์–‘์„ ์ฆ์ง„์‹œ์ผœ ๊ตฌ์ถ• ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ(TA12000, TA24000)์˜ ์–‘์ด ๋™์ผํ•˜๊ฒŒ ๊ตฌ์ถ•๋˜ ์—ˆ์„ ๋•Œ, ์‹œํ—˜ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๊ท ์—ด ๊ฒ€์ถœ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•œ ๊ทธ๋ฆผ์ด ๋‹ค. ๊ทธ๋ฆผ์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด ์›๋ณธ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ(T12000, T24000)์™€ ์ด๋ฏธ์ง€ ๋ณ€ํ™˜ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ์–‘์„ ์ฆ์ง„์‹œํ‚จ ๋ฐ์ดํ„ฐ ์„ธํŠธ(TA12000, TA24000)์— ๋”ฐ๋ฅธ ํ•ฉ์„ฑ๊ณฑ ์‹  ๊ฒฝ๋ง ๋ชจ๋ธ์˜ ๊ท ์—ด ๊ฒ€์ถœ ์„ฑ๋Šฅ์˜ ์ฐจ์ด๋Š” ํฌ๊ฒŒ ๋ฐœ์ƒ๋˜์ง€ ์•Š์Œ์„ ์•Œ ์ˆ˜ ์žˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด๋ฏธ์ง€ ๋ณ€ํ™˜ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ์˜ ์–‘ ์„ ์ฆ์ง„์‹œํ‚ค๋Š” ๋ฐฉ๋ฒ•์€ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์˜ ๊ท ์—ด ๊ฒ€์ถœ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ์œ„ํ•ด ์œ ์šฉํ•˜๊ฒŒ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ํŒ๋‹จ๋œ๋‹ค.

Fig. 8

Accuracy, recall, and precision for the same amount of training data

JKSMI-23-6-38_F8.jpg
Fig. 9

Missing rate, over rate for the same amount of training data

JKSMI-23-6-38_F9.jpg

5. ๊ฒฐ ๋ก 

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง์˜ ํ•™์Šต์— ์‚ฌ์šฉ๋˜๋Š” ์›๋ณธ ํ•™ ์Šต ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ์–‘์˜ ์ฆ์ง„๊ณผ ์ด๋ฏธ์ง€ ๋ณ€ํ™˜์„ ํ†ตํ•ด ํ•™์Šต ๋ฐ์ด ํ„ฐ์˜ ์–‘์„ ์ฆ์ง„์‹œ์ผฐ์„ ๋•Œ์˜ ์„ฑ๋Šฅ์ง€ํ‘œ๋ณ€ํ™”๋ฅผ ํ™•์ธํ•˜์˜€๋‹ค. ํ•™ ์Šต์— ์‚ฌ์šฉ๋˜๋Š” ๋ฐ์ดํ„ฐ์–‘์˜ ์ฆ์ง„์€ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์„ ํ™•์ธํ•˜์˜€๋‹ค. ๊ฐ€์žฅ ๋งŽ์€ ์ˆ˜์˜ ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ์ถ• ๋œ TA96000์„ ํ•™์Šต์— ์‚ฌ์šฉํ•˜์˜€์„ ๊ฒฝ์šฐ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง์˜ ํ‘œ ๋ฉด๊ท ์—ด ๊ฒ€์ถœ ์ •ํ™•๋„, ์ •๋ฐ€๋„, ์žฌํ˜„์œจ์€ 96.6% ์ด์ƒ, ๋ฏธ๊ฒ€์ถœ์œจ, ๊ณผ๊ฒ€์ถœ์œจ์€ 3.4% ์ดํ•˜์˜ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋‚˜ํƒ€๋‚ด์—ˆ๋‹ค.

ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง ํ•™์Šต์— ์žˆ์–ด์„œ ์›๋ณธ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉ ํ•œ ๊ฒฝ์šฐ์™€ ์ด๋ฏธ์ง€ ๋ณ€ํ™˜์„ ํ†ตํ•ด ๋™์ผํ•œ ์–‘์œผ๋กœ ๊ตฌ์ถ•ํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•œ ๊ฒฝ์šฐ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง์˜ ํ‘œ๋ฉด๊ท ์—ด ๊ฒ€์ถœ ์„ฑ๋Šฅ ์ฐจ ์ด๊ฐ€ ํฌ๊ฒŒ ๋ฐœ์ƒ๋˜์ง€ ์•Š์•˜๋‹ค. ๋”ฐ๋ผ์„œ ์ด๋ฏธ์ง€ ๋ณ€ํ™˜ ๋ฐฉ์‹์€ ํ•ฉ์„ฑ ๊ณฑ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์˜ ์„ฑ๋Šฅํ–ฅ์ƒ์„ ์œ„ํ•œ ๋ฐ์ดํ„ฐ์–‘์˜ ์ฆ์ง„ ๋ฐฉ๋ฒ•์œผ ๋กœ ์œ ์šฉํ•˜๊ฒŒ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋‹ค.

๊ฒ€์ถœ ์„ฑ๋Šฅ ํ™•์ธ์„ ํ†ตํ•ด ์•„์ŠคํŒ”ํŠธ ์ฝ˜ํฌ๋ฆฌํŠธ ๋„๋กœํฌ์žฅ ํ‘œ๋ฉด ๊ท ์—ด ๊ฒ€์ถœ ๋ฐฉ์‹์— ์žˆ์–ด ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง์˜ ํ™œ์šฉ ๊ฐ€๋Šฅ์„ฑ์„ ํ™•์ธ ํ•˜์˜€์œผ๋ฉฐ, ํ–ฅํ›„ ์•„์ŠคํŒ”ํŠธ ์ฝ˜ํฌ๋ฆฌํŠธ ๋„๋กœํฌ์žฅ์˜ ํ‘œ๋ฉด๊ฒฐํ•จ์ข…๋ฅ˜ ์— ๋”ฐ๋ฅธ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ๋ฅผ ํ™•๋ณดํ•˜์—ฌ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง์˜ ํ•™์Šต์— ํ™œ์šฉ๋œ๋‹ค๋ฉด, ํ‘œ๋ฉด๊ฒฐํ•จ์˜ ์ข…๋ฅ˜์— ๋”ฐ๋ฅธ ๋ถ„๋ฅ˜ ๋˜ํ•œ ๊ฐ€๋Šฅํ•  ๊ฒƒ์œผ ๋กœ ๊ธฐ๋Œ€๋œ๋‹ค.

๊ฐ์‚ฌ์˜ ๊ธ€

๋ณธ ์—ฐ๊ตฌ๋Š” 2018๋…„ ํ•œ๊ตญ์—ฐ๊ตฌ์žฌ๋‹จ ์ด๊ณต๋ถ„์•ผ๊ธฐ์ดˆ์—ฐ๊ตฌ์ง€์›์‚ฌ ์—…์˜ ์—ฐ๊ตฌ๋น„์ง€์›(2018R1D1A1B07048341)์— ์˜ํ•ด ์ˆ˜ํ–‰๋˜์—ˆ ์Šต๋‹ˆ๋‹ค.

References

1. 
Zhang, L., Yang, F., Zhang, Y.D., Zhu, Y.J. (2016), Road crack detection using deep convolutional neural network, In Proceedings of the 2016 IEEE International Conference on Image Processing (ICIP), Phoenix, AZ, USA, 25โ€“-28 September 2016; pp. 3708-3712.DOI
2. 
FENG, C., Liu, M.-Y., Kao, C.-C., Lee, T.-Y. (2017), Deep Active Learning for Civil Infrastructure Defect Detection and Classification, International Workshop on Computing in Civil Engineering (IWCCE), pp. 298-306.DOI
3. 
Eisenbatch, Markus (2017), How to get pavement distress detection ready for deep learning? A systematic approach, 2017 international joint conference on neural networks (IJCNN). IEEE, pp. 2039-2047.DOI
4. 
Loffe, S., Szegedy, C. (2015), Batch normalization: Acceleration deep network training by redusing internal covariate shift, arXiv preprint arXiv: 1502.03167.Google Search
5. 
Pauly, L., Hogg, D., Fuentes, R., Peel, H. (2017), Deeper networks for pavement crack detection, In Proceedings of the 34th International Symposium on Automation and Robotics in Construction (ISARC). IAARC, pp. 479-485.DOI
6. 
Wu, J. (2017), Introduction to convolutional neural networks. National Key Lab for Novel Software Technology, Nanjing University. China, 5-23.Google Search
7. 
Keskar, N. S., Mudigere, D., Nocedal, J., Smelyanskiy, M., Tang, P. T. P. (2016), On large-batch training for deep learning: Generalization gap and sharp minima, arXiv preprint arXiv; 1609.04836.Google Search
8. 
Ruder, S. (2016), An overview of gradient descent optimization algorithms, arXiv preprint arXiv: 16090.04747.Google Search
9. 
Gopalakrishnan, K. (2018), Deep Learning in Data-Driven Pavement Image Analysis and Automated Distress Detection: A Review. Data 3(3), 28.DOI
10. 
Kim, J. H., Kim, J. R., Moon, H. C. (2008), Development of Pavement Distress Survey System, Journal of Korean Society of Road Engineers (KSRE), pp. 475-484.Google Search
11. 
Choi, Y. S. (2019), Concrete Surface Defect Detection using Convolutional Neural Network, Department of Architectural Engineering Graduate School, Chungbuk National University.Google Search
12. 
Radford, A., Metz, L., Chintala, S. (2015), Unsupervised representation learning with deep convolutional generative adversarial networks, arXiv preprint arXiv:1511.06434.Google Search
13. 
Powers, D. M. (2011), Evaluation: from precision, recall and F-measure to ROC, informedness, markedness and correlation.Google Search
14. 
Bengio, Y., Goodfellow, I. J., Courville, A. (2015), Deep learning, book in preparation for mit press, Disponฤฑvel em http://www. iro. umontreal. ca/bengioy/dlbook.Google Search
15. 
Goodfellow, I., Bengio, Y., Courville, A. (2016), Deep learning, MIT press, pp. 321-362.Google Search
16. 
Hu, J., Shen, L., Sun, G. (2018), Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 7132-7141).DOI
17. 
Rababaah, H., Vrajitoru, D., Wolfer, J. (2005), Asphalt pavement crack classification: a comparison of GA, MLP, and SOM, In Proceedings of Genetic and Evolutionary Computation Conference, Late-Breaking Paper.Google Search
18. 
Koch, C., Brilakis, I (2011), Pothole detection in asphalt pavement images, Advanced Engineering Informatics, 25(3), pp. 507-515.DOI
19. 
Sorncharean, S., Phiphobmongkol, S. (2008), Crack detection on asphalt surface image using enhanced grid cell analysis, In 4th IEEE International Symposium on Electronic Design, Test and Applications (delta 2008), pp. 49-54.DOI