• ๋Œ€ํ•œ์ „๊ธฐํ•™ํšŒ
Mobile QR Code QR CODE : The Transactions of the Korean Institute of Electrical Engineers
  • COPE
  • kcse
  • ํ•œ๊ตญ๊ณผํ•™๊ธฐ์ˆ ๋‹จ์ฒด์ด์—ฐํ•ฉํšŒ
  • ํ•œ๊ตญํ•™์ˆ ์ง€์ธ์šฉ์ƒ‰์ธ
  • Scopus
  • crossref
  • orcid

  1. (Dept. of Electronics Engineering, Seokyeong University, Korea)
  2. (Dept. of Electronics Engineering, Seokyeong University, Korea)



Defects detection, Metal surface, Convolution neural network, Faster R-CNN, YOLOv2

1. ์„œ๋ก 

์ œํ’ˆ์˜ ํ‘œ๋ฉด ๊ฒฐํ•จ์€ ๊ธํž˜, ๊ฐˆ๋ผ์ง, ์ฐํž˜, ํŒจ์ž„, ์–ผ๋ฃฉ ๋“ฑ์„ ํฌํ•จํ•œ๋‹ค. ํ‘œ๋ฉด ์žฌ์งˆ์˜ ํŠน์„ฑ์— ๋”ฐ๋ผ ์กฐ๋ช… ๋ฐ˜์‚ฌ๊ฐ€ ์‹ฌํ•˜๊ฑฐ๋‚˜, ๊ฐ€๊ณต๊ณผ์ •์—์„œ ๋ฐœ์ƒํ•˜๋Š” ๋ถˆ๊ท ์ผํ•œ ๋ฌด๋Šฌ ๋“ฑ์ด ์กด์žฌํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ •์ƒ ์˜์—ญ๊ณผ ๊ฒฐํ•จ ์˜์—ญ์˜ ๊ตฌ๋ณ„์ด ๋งค์šฐ ์–ด๋ ต๋‹ค. ์‚ฐ์—…์šฉ ํ‘œ๋ฉด ๊ฒฐํ•จ ๊ฒ€์‚ฌ๋Š” ๋†’์€ ์ •ํ™•๋„์™€ ๋น ๋ฅธ ์ฒ˜๋ฆฌ๊ฐ€ ํ•„์š”ํ•˜๋ฉฐ, ๊ฒฐํ•จ์˜ ์œ ๋ฌด ์™ธ์— ๊ฒ€์ถœ ์œ„์น˜์— ๋Œ€ํ•œ ํ‘œ์‹œ๊ฐ€ ํ•„์š”ํ•œ ๊ฒฝ์šฐ๊ฐ€ ๋งŽ๋‹ค.

ํ•œํŽธ, ์ ์šฉ ๊ธฐ๋ฒ•์˜ ์ธก๋ฉด์„ ๋ณด๋ฉด ์ „ํ†ต์ ์ธ ํ•„ํ„ฐ ๊ธฐ๋ฐ˜์˜ ์ ‘๊ทผ๋ฒ•(1), ํ•„ํ„ฐ์™€ SVM๋“ฑ์˜ ๊ธฐ๊ณ„ํ•™์Šต์„ ๊ฒฐํ•ฉํ•œ ์ ‘๊ทผ๋ฒ•(2), ๊ทธ๋ฆฌ๊ณ  CNN(Convolutional Neural Network)(3) ๋“ฑ์„ ์ด์šฉํ•œ ๋”ฅ๋Ÿฌ๋‹ ์ ‘๊ทผ๋ฒ•์ด ์žˆ๋‹ค. ๋‚œ์ด๋„๊ฐ€ ๋†’์€ ๊ฒฐํ•จ ๊ฒ€์‚ฌ์—๋Š” ํ•„ํ„ฐ ๋˜๋Š” ๊ฒ€์ถœ์ž ๊ธฐ๋ฐ˜์˜ ์ ‘๊ทผ์ด ์„ฑ๋Šฅ์˜ ํ•œ๊ณ„๋ฅผ ๋ณด์ด๊ณ  ์žˆ๋‹ค. ์กฐ๋ช…์˜ ๋ถˆ๊ท ์ผํ•œ ๋ฐ˜์‚ฌ๋‚˜ ํ‘œ๋ฉด์ด ๊ท ์ผํ•˜์ง€ ์•Š์€ ๊ฒฝ์šฐ, ๊ฒฐํ•จ๊ณผ ์ฃผ๋ณ€์˜ ๊ตฌ๋ถ„์ด ๋ช…ํ™•์น˜ ์•Š๊ธฐ ๋•Œ๋ฌธ์— ๊ฒฐํ•จ์„ ๊ฒ€์ถœํ•˜๊ธฐ๊ฐ€ ๋”์šฑ ์–ด๋ ต๋‹ค. ๋˜ํ•œ ํ•ด์ƒ๋„๊ฐ€ ๋†’์€ ์˜์ƒ์—์„œ ์ž‘์€ ๊ฒฐํ•จ์„ ๋ฐœ๊ฒฌ์„ ํ•ด์•ผ ํ•˜๋Š” ๊ฒฝ์šฐ๋„ ์ข…์ข… ์žˆ๋‹ค.

๋”ฐ๋ผ์„œ ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•˜์—ฌ ์ตœ๊ทผ ๋ฌผ์ฒด ์ธ์‹์— ํƒ์›”ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ด๊ณ  ์žˆ๊ณ  ๊ณ„์†์ ์ธ ๋ฐœ์ „์ด ์ด๋ฃจ์–ด์ง€๊ณ  ์žˆ๋Š” ๋”ฅ๋Ÿฌ๋‹ ๊ธฐ๋ฒ•์„ ๊ฒฐํ•จ ๊ฒ€์‚ฌ์— ์ ์šฉํ•˜๋Š” ์—ฐ๊ตฌ๋“ค์ด ๋Š˜๊ณ  ์žˆ๋‹ค(4-6). ํŠนํžˆ, CNN์€ ์ž๋™์ ์ธ ํŠน์ง• ์ถ”์ถœ ๊ธฐ๋Šฅ์ด ๊ฐ•๋ ฅํ•˜์—ฌ, ํŠน์ง• ์ถ”์ถœ์„ ์œ„ํ•œ ์ „์ฒ˜๋ฆฌ ํ•„ํ„ฐ๊ฐ€ ํ•„์š”ํ•˜์ง€ ์•Š๋‹ค๊ณ  ์•Œ๋ ค์ ธ ์žˆ๋‹ค. ๊ธฐ์กด์˜ ๋”ฅ๋Ÿฌ๋‹ ์ ‘๊ทผ๋ฒ•๋“ค์€ ์ฃผ๋กœ ๋ฌผ์ฒด์— ๋Œ€ํ•œ ๊ฒ€์ถœ ์‹œ๋„๊ฐ€ ๋Œ€๋ถ€๋ถ„์ด๊ธฐ ๋•Œ๋ฌธ์—, ๊ฒฐํ•จ์— ๋Œ€ํ•ด์„œ ์ ์šฉ ์‹œ ๊ฐ์ฒด์™€์˜ ํŠน์„ฑ ์ฐจ์ด๋ฅผ ๋ถ„์„ํ•˜๊ณ  ์ด๋ฅผ ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ๋ฐ˜์˜ํ•˜๋Š” ๊ฒƒ์ด ํ•„์š”ํ•˜๋‹ค.

์„ ํ–‰ ์—ฐ๊ตฌ์—์„œ VOV(7) ํ•„ํ„ฐ์™€ CNN์„ ๊ฒฐํ•ฉํ•˜์—ฌ ํ‘œ๋ฉด ๊ฒฐํ•จ์— ๋Œ€ํ•œ ๊ฒ€์ถœ์„ ์‹œ๋„ํ•˜์˜€๋‹ค(6). ์ด ๋ฐฉ๋ฒ•์€ ROI์˜ ์ˆ˜๊ฐ€ ๋น„๊ต์  ์ ์–ด์„œ ์ฒ˜๋ฆฌ ์†๋„ ๋ฉด์—์„œ ์žฅ์ ์„ ๊ฐ€์ง€๋‚˜, ๊ฒ€์ถœ์„ฑ๋Šฅ์ด ๋งŒ์กฑ์Šค๋Ÿฝ์ง€ ์•Š์•˜๋‹ค. ํ›„์† ์—ฐ๊ตฌ๋กœ R-CNN(8,9) ๊ธฐ๋ฐ˜ ์ ‘๊ทผ๋ฒ• ์ค‘ ๊ฐ€์žฅ ์ตœ์‹  ๋ฐฉ๋ฒ•์ธ Faster R-CNN(10)์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฒ€์ถœ ์„ฑ๋Šฅ๊ณผ ์†๋„๋ฅผ ํ–ฅ์ƒ์‹œ์ผฐ๋‹ค(11). ๋˜ํ•œ, YOLOv2(12) ๊ธฐ๋ฒ•์„ ์ ์šฉํ•˜์—ฌ Faster R-CNN ์ ‘๊ทผ๋ฒ•๊ณผ ๋น„๊ตํ•œ ์—ฐ๊ตฌ๋„ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค(13). ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ์‚ฌ์ „ ์—ฐ๊ตฌ๋ฅผ ํ™•์žฅํ•˜์—ฌ, ์‚ฐ์—…์šฉ ๊ฒฐํ•จ ๊ฒ€์‚ฌ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด์„œ CNN ๊ธฐ๋ฐ˜์˜ ์ƒ๊ธฐ์˜ ๊ฒ€์ถœ ๊ธฐ๋ฒ•๋“ค๊ฐ„์˜ ๋น„๊ต ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•˜๊ณ  ์„ฑ๋Šฅ์„ ๋ถ„์„ํ•œ๋‹ค.

2. ๊ฒฐํ•จ ๊ฒ€์‚ฌ ๊ตฌ๋ถ„ ๋ฐ ์œ„์น˜๊ฒ€์ถœ

2.1 ๊ฒฐํ•จ ๊ฒ€์ถœ ๋ฌธ์ œ์˜ ํŠน์„ฑ

์‚ฐ์—…์šฉ ๋ถ€ํ’ˆ ๋“ฑ์—์„œ ํ‘œ๋ฉด ๊ฒฐํ•จ์˜ ์ข…๋ฅ˜๋Š” ํฌ๋ž™, ์Šคํฌ๋ž˜์น˜, ํ ์ง‘ ๋“ฑ์„ ํฌํ•จํ•œ๋‹ค. ํ‘œ๋ฉด์ด ๊ท ์ผํ•˜์ง€ ์•Š๊ฑฐ๋‚˜ ๋งค๋„๋Ÿฝ์ง€ ์•Š์€ ๊ฒฝ์šฐ, ํ‘œ๋ฉด ๋ฐฐ๊ฒฝ๊ณผ ๊ฒฐํ•จ์˜ ๊ตฌ๋ณ„์ด ๋งค์šฐ ์–ด๋ ต๋‹ค. ๊ทธ๋ฆผ 1์€ ๊ด€๋ จ ๊ธˆ์† ๋ถ€ํ’ˆ์˜ ํ‘œ๋ฉด ์˜ˆ์ด๋‹ค. ์กฐ๋ช…์˜ ๋ถˆ๊ท ์ผ๋กœ ์ธํ•œ ์ฃผ๋ณ€์˜ ๋ฐฐ๊ฒฝ ์˜์ƒ์ด ๊ฒฐํ•จ๊ณผ ์ž˜ ๊ตฌ๋ณ„๋˜์ง€ ์•Š๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ๋‹ค. ์žฌ์งˆ ํŠน์„ฑ๊ณผ ๊ฐ€๊ณต ๊ณผ์ •์—์„œ ํฌํ•จ๋˜๋Š” ๋ฐฐ๊ฒฝ ๋ฌด๋Šฌ๋“ค๊ณผ ์กฐ๋ช…์˜ ๊ฒฐํ•ฉ์€ ๊ฒฐํ•จ์˜ ๊ฒ€์ถœ์„ ๋”์šฑ ์–ด๋ ต๊ฒŒ ํ•œ๋‹ค.

2.2 ๊ฒฐํ•จ์˜ ๋ถ„๋ฅ˜ ๋ฐ ์œ„์น˜ ๊ฒ€์ถœ

์‚ฐ์—…์šฉ ๋ถ€ํ’ˆ์— ๋Œ€ํ•œ ๊ฒฐํ•จ ๊ฒ€์‚ฌ๋Š” ๊ฒฐํ•จ ์œ„์น˜๊ฐ€ ๊ณ ์ •๋˜์–ด ์žˆ๊ณ , ๊ฒฐํ•จ์˜ ์—ฌ๋ถ€๋‚˜ ์ข…๋ฅ˜๋งŒ์„ ํŒ๋‹จํ•˜๋Š” ๊ฒฐํ•จ ๋ถ„๋ฅ˜์˜ ๋ฌธ์ œ์™€ ๊ฒฐํ•จ์˜ ์œ„์น˜๊นŒ์ง€ ํŒ๋ณ„ํ•ด์•ผ ํ•˜๋Š” ๊ฒฐํ•จ ๊ฒ€์ถœ ๋ฌธ์ œ๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์žˆ๋‹ค. ๋”์šฑ์ด ๋Œ€์ƒ ๋ถ€ํ’ˆ์ด ํ‰๋ฉด์ด ์•„๋‹Œ ๊ฒฝ์šฐ, ์บ ์ด๋‚˜ ๋ถ€ํ’ˆ์„ ์ด๋™์‹œ์ผœ ๋‹ค์ˆ˜์˜ ์˜์ƒ์„ ์–ป์–ด์•ผํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ด ๊ณผ์ •์—์„œ ์กฐ๋ช…์˜ ๋‚œ๋ฐ˜์‚ฌ ์ดˆ์ ์˜ ํ๋ฆผ ํ˜„์ƒ์ด ๋‚˜ํƒ€๋‚˜ ๊ฒ€์ถœ ์กฐ๊ฑด์„ ์–ด๋ ต๊ฒŒ ๋งŒ๋“ค ์ˆ˜ ์žˆ๋‹ค. ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ํ›„์ž๋ฅผ ๋‹ค๋ฃฌ๋‹ค. ์•„๋ž˜ ๊ทธ๋ฆผ 2๋Š” ๊ฒ€์‚ฌ ๋Œ€์ƒ ๋ถ€ํ’ˆ์„ ๊ณ ํ•ด์ƒ๋„ ์บ ์„ ์ด๋™์‹œํ‚ค๋ฉฐ ๊ธˆ์† ๋ถ€ํ’ˆ์˜ ํ‘œ๋ฉด์„ ์ดฌ์˜ํ•˜๊ณ  ๊ฒฐํ•จ์„ ๊ฒ€์ถœํ•˜๋Š” ์‹œ์Šคํ…œ์˜ ๊ฐœ๋…๋„์ด๋‹ค.

๊ทธ๋ฆผ. 1. ํ‘œ๋ฉด ๊ฒฐํ•จ์˜ ์˜ˆ

Fig. 1. Exmaples of surface defects

../../Resources/kiee/KIEE.2018.67.7.865/fig1.png

3. CNN ๊ธฐ๋ฐ˜ ๊ฒฐํ•จ ๊ฒ€์ถœ ๊ธฐ๋ฒ•

3.1 CNN(Convolutional Neural Network)

CNN(3)์€ 1989๋…„ LeCun์ด ๋ฐœํ‘œํ•œ ๊ธฐ๋ฒ•์œผ๋กœ, ์˜์ƒ์ธ์‹์— ์ฃผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๋”ฅ๋Ÿฌ๋‹ ๊ธฐ๋ฒ•์ด๋‹ค. ๊ธฐ์กด์˜ ์ „ํ†ต์  ๋ฌผ์ฒด ๋ถ„๋ฅ˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜๋ณด๋‹ค 10%์ด์ƒ ์ข‹์€ ์„ฑ๋Šฅ์„ ๋ณด์ด๋ฉด์„œ ILSVRC-2012(8)๋ถ€ํ„ฐ ์ฃผ๋ฅ˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ ์‚ฌ์šฉ๋˜๊ณ  ์žˆ๋‹ค.

CNN์€ ํŠน์ง•๋“ค์„ ์‚ฌ์ „์— ์ถ”์ถœํ•˜์ง€ ์•Š๊ณ , ์ž…๋ ฅ ์ธต์— ์˜์ƒ ๋ฐ์ดํ„ฐ๋ฅผ ์ž…๋ ฅํ•˜์—ฌ ํŠน์ง•์ด ์ž๋™ ์ถ”์ถœ๋˜๋Š” ํŠน์ง•์„ ๊ฐ€์ง€๊ณ  ์žˆ๋‹ค. ์ด๋ฏธ์ง€๋ฅผ ์ž…๋ ฅ๋ฐ›์•„ ํŠน์ง•์„ ์ž๋™์œผ๋กœ ์ถ”์ถœํ•œ๋‹ค.

์ปจ๋ณผ๋ฃจ์…˜(convolution) ์ธต๊ณผ ํ’€๋ง(pooling) ์ธต์ด ๊ต๋Œ€๋กœ ๋ฐ˜๋ณต๋˜๋Š” ๊ตฌ์กฐ๋ฅผ ์ด๋ฃฌ๋‹ค. ์ปจ๋ณผ๋ฃจ์…˜(ํ•ฉ์„ฑ๊ณฑ) ์—ฐ์‚ฐ์€ ์ด๋ฏธ์ง€๋ฅผ ํ๋ฆฟํ•˜๊ฒŒ ํ•˜๊ฑฐ๋‚˜ ๋ชจ์„œ๋ฆฌ๋‚˜ ์„ ์„ ๊ฐ•์กฐํ•˜๋Š” ํšจ๊ณผ๋ฅผ ์ค€๋‹ค. ํ’€๋ง ์—ฐ์‚ฐ์€ ์ด๋ฏธ์ง€ ๋‚ด ํŠน์ง•์˜ ์ž‘์€ ์œ„์น˜๋ณ€ํ™”์— ๋Œ€ํ•œ ๋ถˆ๋ณ€์„ฑ์„ ๋ถ€์—ฌํ•œ๋‹ค. ์ด ๋‘ ๊ฐ€์ง€ ์ธต์„ ๊ต๋Œ€๋กœ ๊ฑฐ์น˜๋ฉด์„œ ์ด๋ฏธ์ง€์˜ ํŠน์ง•์„ ์ถ”์ถœํ•˜์—ฌ ์ตœ์ข…์ ์œผ๋กœ ํŠน์ง• ๋งต์„ ๋งŒ๋“ค๊ณ , ์ด๋ฅผ ์™„์ „์—ฐ๊ฒฐ(Fully connected) ์ธต์— ์ „๋‹ฌํ•˜์—ฌ ์ด๋ฏธ์ง€๋ฅผ ๋ถ„๋ฅ˜ํ•œ๋‹ค. ๊ทธ๋ฆผ 3์€ ์‚ฌ์ „ ์—ฐ๊ตฌ (6)์—์„œ ์‚ฌ์šฉ๋œ CNN ๋ชจ๋ธ ๊ตฌ์กฐ์˜ ์˜ˆ์ด๋ฉฐ, ๋งต๊ณผ ์ปค๋„์˜ ํฌ๊ธฐ๋Š” ๊ฒฐํ•จ์— ์ ํ•ฉํ•˜๊ฒŒ ์„ค์ •๋˜์–ด ์žˆ๋‹ค.

๊ทธ๋ฆผ. 2. ๊ฒ€์‚ฌ ์‹œ์Šคํ…œ ๊ตฌ์„ฑ๋„

Fig. 2. Construction of inspection system

../../Resources/kiee/KIEE.2018.67.7.865/fig2.png

๊ทธ๋ฆผ. 3. CNN ๋ชจ๋ธ ๊ตฌ์กฐ

Fig. 3. CNN model structure

../../Resources/kiee/KIEE.2018.67.7.865/fig3.png

CNN์œผ๋กœ ์ง์ ‘ ํ•™์Šต๋งŒ์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ฐฉ๋ฒ•์€ ์‚ฌ๋žŒ, ๊ณ ์–‘์ด, ๋ฐฐ๋“ฑ์˜ ๋ฌผ์ฒด ์ธ์‹์— ๋งŽ์ด ์“ฐ์ด๊ณ , ๋›ฐ์–ด๋‚œ ์ธ์‹๋ฅ ์„ ๋ณด์ด๊ณ  ์žˆ๋‹ค. ๊ทธ ์ด์œ ๋Š” ๋Œ€์ƒ ๋ฌผ์ฒด์˜ ํŠน์ง•์ด ์ผ๊ด€๋˜๊ณ , ๋‹ค๋ฅธ ๋ฌผ์ฒด์™€ ๊ตฌ๋ณ„๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋น„๊ต์  ํฌ๊ธฐ๋„ ์ปค์„œ ์ „์ฒ˜๋ฆฌ ์—†์ด ํ•™์Šต์ด ๊ฐ€๋Šฅํ•˜๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์‚ฐ์—…์šฉ ๋ถ€ํ’ˆ์— ๋Œ€ํ•œ ๊ฒฐํ•จ์˜ ๊ฒฝ์šฐ๋Š” ํ˜•ํƒœ์— ๋Œ€ํ•œ ์ผ๊ด€๋œ ํŠน์ง•์ด ์—†๊ณ , ํฌ๊ธฐ๋„ ์ž‘์€ ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์•„์„œ ํ•™์Šต๋งŒ์œผ๋กœ ํŠน์ง•์„ ์ž๋™์ถ”์ถœํ•˜๊ธฐ๊ฐ€ ์–ด๋ ต๋‹ค.

๋˜ํ•œ, ๊ฒฐํ•จ์˜ ์œ„์น˜๋ฅผ ํ‘œ์‹œํ•ด์•ผ ํ•˜๋Š” ๊ฒฝ์šฐ, CNN ๋งŒ์˜ ํ•™์Šต์œผ๋กœ๋Š” ์–ด๋ ค์šฐ๋ฉฐ, R-CNN๊ณผ ๊ฐ™์ด ์„ ํƒ์  ํƒ์ƒ‰์„ ํ†ตํ•ด ํ›„๋ณด์˜์—ญ ๊ตฐ๋“ค์„ ๊ฒ€์ถœํ•  ์ˆ˜ ์žˆ์–ด์•ผ ๊ฐ€๋Šฅํ•˜๋‹ค. ๊ทธ๋Ÿฌ๋‚˜, ์˜์ƒ ๋‹น ์ˆ˜๋ฐฑ ๊ฐœ์—์„œ ์ˆ˜์ฒœ ๊ฐœ์˜ ํ›„๋ณด ์˜์—ญ๋“ค์„ ๊ฒ€์ถœํ•˜๋ฏ€๋กœ ๋ถ€๊ฐ€์ ์ธ ์—ฐ์‚ฐ๋Ÿ‰์ด ์ฆ๊ฐ€๋˜๋Š” ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•œ๋‹ค.

3.2 Faster R-CNN

Faster R-CNN์€ CNN์„ ํ™•์žฅํ•œ ๊ธฐ๋ฒ•์œผ๋กœ ์ด๋ฏธ์ง€ ๋‚ด์˜ ๋ฌผ์ฒด ๊ฒ€์ถœ ์‹œ๊ฐ„์˜ ๋‹จ์ถ• ๋ฐ ๊ฒ€์ถœ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ์ด๋ฃจ์—ˆ๋‹ค. ๋ฌผ์ฒด ๊ฒ€์ถœ์€ ๊ฐ์ฒด์˜ ํด๋ž˜์Šค ๋ถ„๋ฅ˜ ๋ฟ ์•„๋‹ˆ๋ผ ์ด๋ฏธ์ง€ ์ƒ์—์„œ ๋ฌผ์ฒด์˜ ์œ„์น˜, ๋„“์ด ๋ฐ ํญ์„ ๊ฒฝ๊ณ„์ƒ์ž๋กœ ๋‚˜ํƒ€๋‚ธ๋‹ค. CNN ๊ธฐ๋ฐ˜์˜ ๋ฌผ์ฒด ๊ฒ€์ถœ ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ R-CNN, Fast R-CNN์„ ๊ฑฐ์น˜๋ฉฐ ๊ฒ€์ถœ ์†๋„์™€ ์„ฑ๋Šฅ์ด ๊ฐœ์„ ๋˜์–ด ์™”๋‹ค(8,9). Fast R-CNN์€ Selective Search ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‚ฌ์šฉํ•˜์—ฌ ํ›„๋ณด ์˜์—ญ(ROI)์„ ์ œ์•ˆํ•˜๊ณ  ์ด๋ฅผ ๊ฒฝ๊ณ„์ƒ์ž(bounding box) ํšŒ๊ท€(regressors)๋กœ ํ•™์Šต์‹œ์ผฐ๋Š”๋ฐ, Faster R-CNN(10)์—์„œ๋Š” ์ด๋ฅผ ๋ณด๋‹ค ์ฒด๊ณ„ํ™”ํ•˜์—ฌ, ๋‹ค์–‘ํ•œ ํฌ๊ธฐ์˜ ์•ต์ปค(anchor)๋ฅผ ์ ์šฉํ•œ RPN(Region Proposal Network)์„ ์ œ์•ˆํ•˜์˜€๋‹ค.

Faster R-CNN ๋‚ด์˜ ์ปจ๋ณผ๋ฃจ์…˜ ๋„คํŠธ์›Œํฌ๋Š” ์ผ๋ฐ˜์ ์ธ CNN๊ณผ ๊ฐ™์ด ์ปจ๋ณผ๋ฃจ์…˜ ์ธต๊ณผ ํ’€๋ง ์ธต์œผ๋กœ ์ด๋ฃจ์–ด์ ธ ์žˆ๊ณ , ์—ฌ๊ธฐ์„œ ์ถœ๋ ฅ๋˜๋Š” ํŠน์ง•๋งต์€ RPN๊ณผ ์™„์ „์—ฐ๊ฒฐ ์ธต์œผ๋กœ ์—ฐ๊ฒฐ๋œ๋‹ค. RPN์€ ํŠน์ง•๋งต์— ์Šฌ๋ผ์ด๋”ฉ ์œˆ๋„์šฐ ๋ฐฉ์‹์œผ๋กœ ์•ต์ปค(anchor) ๋ฐ•์Šค๋ฅผ ์ ์šฉํ•œ๋‹ค. ์•ต์ปค ๋ฐ•์Šค๋Š” ํฌ๊ธฐ์™€ ๋น„์œจ์„ ๋ณ€ํ™”์‹œํ‚จ ๋‹ค์–‘ํ•œ ํฌ๊ธฐ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ๊ณ , ์ด๋“ค์„ ์ ์šฉ์‹œ์ผœ ๊ฒ€์ถœํ•˜๊ณ ์ž ํ•˜๋Š” ๊ฐ์ฒด๊ฐ€ ์กด์žฌํ•  ๋งŒํ•œ ๋‹ค์–‘ํ•œ ์˜์—ญ์„ ์ œ์•ˆํ•œ๋‹ค. ์ œ์•ˆ๋œ ์˜์—ญ ์ค‘ IOU๋ฅผ ๊ณ„์‚ฐํ•˜์—ฌ ์ผ์ • ์ˆ˜์น˜(์˜ˆ๋กœ 0.7) ์ด์ƒ์„ ์ตœ์ข… ๊ฒ€์ถœ ์˜์—ญ์œผ๋กœ ๊ฒฐ์ •ํ•œ๋‹ค. ๊ทธ๋ฆผ 4๋Š” Faster R-CNN์˜ ๊ตฌ์กฐ์ด๋‹ค.

๊ทธ๋ฆผ. 4. Faster R-CNN ๊ตฌ์กฐ

Fig. 4. Faster R-CNN structure

../../Resources/kiee/KIEE.2018.67.7.865/fig4.png

๊ทธ๋ฆผ. 5. YOLOv2 ๊ตฌ์กฐ

Fig. 5. YOLOv2 structure

../../Resources/kiee/KIEE.2018.67.7.865/fig5.png

3.3 YOLO v2

YOLOv2(14)๋Š” ์‹ค์‹œ๊ฐ„ ๊ฐ์ฒด ๊ฒ€์ถœ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” ํ•™์Šต ๋ฐ ๊ฒ€์ถœ ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ ๊ธฐ์กด์˜ YOLO(12) ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๊ฐœ์„ ํ•œ ๊ธฐ๋ฒ•์ด๋‹ค. YOLOv2๋Š” ์ž…๋ ฅ ์ด๋ฏธ์ง€๋ฅผ S ร— S ์˜ ๊ทธ๋ฆฌ๋“œ ์…€๋กœ ๋‚˜๋ˆˆ ํ›„ ๊ฐ ๊ทธ๋ฆฌ๋“œ ์…€๋งˆ๋‹ค 5๊ฐœ์˜ ์•ต์ปค ๋ฐ•์Šค๋ฅผ ์ ์šฉํ•˜์—ฌ ๊ทธ๋ฆฌ๋“œ ์…€ ๋‚ด ๊ฐ์ฒด์˜ ์กด์žฌ ํ™•๋ฅ , ๊ฐ์ฒด์— ๋Œ€ํ•œ ํด๋ž˜์Šค์˜ ํ™•๋ฅ , ๊ฐ์ฒด์˜ ์ค‘์‹ฌ์ขŒํ‘œ์™€ ๊ฐ์ฒด์˜ ๋„ˆ๋น„ ๋ฐ ๋†’์ด๋ฅผ ์ถ”์ •ํ•œ๋‹ค. Faster R-CNN์ด RPN์œผ๋กœ ๊ฐ์ฒด์˜ ์˜์—ญ์„ ์ œ์•ˆํ•˜๋Š” ๊ฒƒ๊ณผ ๋‹ฌ๋ฆฌ, YOLOv2๋Š” ๊ฐ„๋‹จํ•œ ๊ทธ๋ฆฌ๋“œ ์…€์„ ๋ฐ”ํƒ•์œผ๋กœ ๊ฐ์ฒด์˜ ์˜์—ญ์„ ์ œ์•ˆํ•œ๋‹ค. ์™„์ „์—ฐ๊ฒฐ๊ณ„์ธต ๋Œ€์‹  ์ „์—ญํ‰๊ท ํ’€๋ง(global average pooling)์„ ์‚ฌ์šฉํ•˜์—ฌ ํด๋ž˜์Šค๋ฅผ ๋ถ„๋ฅ˜ํ•˜๊ธฐ ๋•Œ๋ฌธ์— Faster R-CNN๊ธฐ๋ฒ•์— ๋น„ํ•ด ๋น ๋ฅธ ๊ฒ€์ถœ์ด ๊ฐ€๋Šฅํ•˜๋‹ค. Faster R-CNN์ด 9๊ฐœ์˜ ์•ต์ปค๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋น„ํ•ด YOLOv2๋Š” 5๊ฐœ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ์ ๋„ ์ž‘์šฉํ•œ๋‹ค.

YOLOv2๋Š” ๋‹ค์ค‘ ์Šค์ผ€์ผ์— ๊ฐ•์ธํ•˜๊ฒŒ ๋Œ€์ฒ˜ํ•˜๊ธฐ ์œ„ํ•ด ๋งˆ์ง€๋ง‰ ํ’€๋ง์ธต์„ ๊ฑฐ์น˜๊ธฐ ์ „์˜ ํŠน์ง•๋งต(feature map)์—์„œ๋„ bbox๋ฅผ ์ถ”์ถœํ•˜์—ฌ ์‚ฌ์šฉํ•œ๋‹ค. ์•ž์ชฝ ํŠน์ง• ๋งต์—์„œ ์ถ”์ถœ๋œ bbox๋Š” ์ž‘์€ ๋ฌผ์ฒด, ํ’€๋ง ์—ฐ์‚ฐํ›„์˜ ํŠน์ง• ๋งต์—์„œ ์ถ”์ถœ๋œ bbox๋Š” ๋ณด๋‹ค ํฐ ๋ฌผ์ฒด์˜ ํŠน์ง•์„ ํฌํ•จํ•œ๋‹ค. ๋˜ํ•œ ํ•™์Šตํ•  ๋•Œ ์ผ์ • ๋ฐฐ์น˜๋งˆ๋‹ค ์ž…๋ ฅ ์ด๋ฏธ์ง€์˜ ํฌ๊ธฐ๋ฅผ ๋žœ๋คํ•œ ์‚ฌ์ด์ฆˆ๋กœ ์กฐ์ •ํ•˜์—ฌ ํ•™์Šตํ•œ๋‹ค. ๊ทธ๋ฆผ 5์— YOLOv2์˜ ๊ตฌ์กฐ๊ฐ€ ๋‚˜์™€ ์žˆ๋‹ค.

4. ์‹คํ—˜ ๋ฐ ๊ฒฐ๊ณผ ๋ถ„์„

๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ๊ธˆ์†์œผ๋กœ ๋œ ๊ธˆ์† ๋ถ€ํ’ˆ์˜ ํ‘œ๋ฉด์— ๋Œ€ํ•ด Faster R-CNN๊ณผ YOLO v2 ๊ธฐ๋ฐ˜์˜ ๊ฒฐํ•จ ๊ฒ€์ถœ ๋ฐฉ๋ฒ•์„ ์‹คํ—˜ํ•˜๊ณ  ๋น„๊ตํ•œ๋‹ค. ๋”๋ถˆ์–ด ์‚ฌ์ „ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์ธ VOV ํ•„ํ„ฐ์™€ CNN์„ ๊ฒฐํ•ฉํ•œ ๊ฒ€์ถœ ๋ฐฉ๋ฒ•์˜ ๊ฒฐ๊ณผ๋„ ํ•จ๊ป˜ ์ •๋ฆฌํ•œ๋‹ค.

์‹คํ—˜์— ์‚ฌ์šฉํ•œ ๊ธˆ์† ๋ถ€ํ’ˆ ํ‘œ๋ฉด์˜ ๊ฒฐํ•จ ์ข…๋ฅ˜๋Š” ๊ฐˆ๋ผ์ง, ๊ธํž˜, ์ฐํž˜, ์˜ค์—ผ ๋“ฑ์„ ํฌํ•จํ•œ๋‹ค. ํ‘œ๋ฉด์˜ ์งˆ๊ฐ์ด ์•ฝ๊ฐ„ ์˜คํ†จ๋„ํ†จํ•˜๊ฑฐ๋‚˜ ๋ถˆ๊ท ์ผํ•œ ๋Š๋‚Œ์ด ์žˆ์–ด์„œ ์ดฌ์˜ ์‹œ ๋ถˆ๊ท ์ผํ•œ ๋ฌด๋Šฌ๊ฐ€ ๋‚˜ํƒ€๋‚œ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฐ๊ฒฝ ๋ฌด๋Šฌ๊ฐ€ ๊ฒฐํ•จ๊ณผ ์œ ์‚ฌํ•˜์—ฌ ์ •์ƒ ์˜์—ญ๊ณผ ๊ฒฐํ•จ ์˜์—ญ์˜ ๊ตฌ๋ณ„์„ ์–ด๋ ต๊ฒŒ ํ•œ๋‹ค.

๋˜ํ•œ, ์ดฌ์˜ ๋•Œ ์ƒ๊ธด ์กฐ๋ช… ๋ฐ˜์‚ฌ๊ฐ€ ๊ฒฐํ•จ์„ ์™œ๊ณกํ•˜๊ฑฐ๋‚˜, ์ •์ƒ ์˜์—ญ๋„ ๋ฐ๊ธฐ์— ๋”ฐ๋ผ ๊ฒฐํ•จ๊ณผ ์œ ์‚ฌํ•˜๊ฒŒ ๋ณด์ด๋Š” ๊ฒฝ์šฐ๋„ ์žˆ๋‹ค. ๋˜ํ•œ ์ œํ’ˆ์˜ ๊ตด๊ณก์ง„ ๋ถ€๋ถ„์œผ๋กœ ์ธํ•˜์—ฌ ์ƒ๊ธฐ๋Š” ์ดˆ์ ์ด ๋งž์ง€ ์•Š๋Š” ๋ถ€๋ถ„๊ณผ ๋จผ์ง€ ๋“ฑ ์žก์Œ์œผ๋กœ ์ธํ•ด ๊ฒฐํ•จ๊ณผ ํ˜ผ๋™๋  ์ˆ˜ ์žˆ์–ด ๊ฒฐํ•จ ๊ฒ€์ถœ์„ ์–ด๋ ต๊ฒŒ ํ•œ๋‹ค. ์ด๋Ÿฌํ•œ ์š”์ธ๋“ค์ด ๊ฒฐํ•จ ๊ฒ€์ถœ์„ ๋”์šฑ ์–ด๋ ต๊ฒŒ ํ•œ๋‹ค. ๋˜ํ•œ, ๊ฒ€์‚ฌ ๋Œ€์ƒ ์ œํ’ˆ์˜ ์œ„์น˜๊ฐ€ ๊ท ์ผํ•˜์ง€ ์•Š์•„์„œ, ๊ฒ€์‚ฌ ์˜์—ญ์ด ์ „ํ›„์ขŒ์šฐ๋กœ ์ด๋™๋˜๊ธฐ ๋•Œ๋ฌธ์— ๊ณ ์ •๋œ ์Šฌ๋ผ์ด๋”ฉ ์œˆ๋„์šฐ ๋ฐฉ์‹์˜ ๊ฒ€์‚ฌ ๋ฐฉ์‹์€ ์ ํ•ฉํ•˜์ง€ ์•Š์œผ๋ฉฐ ๋ณธ ๋…ผ๋ฌธ์˜ ํŠน์ง•๋งต ๋ฐฉ์‹์˜ ์ ‘๊ทผ์ด ํšจ์œจ์ ์ด๋‹ค.

์‹คํ—˜์— ์‚ฌ์šฉ๋œ ๋ฐ์ดํ„ฐ๋Š” ํ‘œ 1๊ณผ ๊ฐ™๋‹ค. VOV+CNN ๋ฐฉ๋ฒ•์€ ํ•™์Šต์‹œ ๋ถ„๋ฅ˜ ๊ฐ€๋Šฅํ•œ ๋ถ€๋ถ„์ ์ธ ์กฐ๊ฐ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์˜€๊ณ , Faster R-CNN, YOLOv2์—์„œ๋Š” ์ „์ฒด ์ด๋ฏธ์ง€๋ฅผ ์‚ฌ์šฉํ•˜์˜€๋‹ค. ๋˜ํ•œ ์œ„์น˜ ๊ฒ€์ถœ์ด ํ•„์š”ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๊ฒฐํ•จ์ด ์žˆ๋Š” ๋ฐ์ดํ„ฐ๋งŒ์„ ์‚ฌ์šฉํ•ด์•ผ ํ•™์Šต์ด ๊ฐ€๋Šฅํ•˜๋‹ค. ๋”ฐ๋ผ์„œ ๋น„๊ฒฐํ•จ ๋ฐ์ดํ„ฐ๋Š” ํ•™์Šต์— ์‚ฌ์šฉํ•˜์ง€ ์•Š์•˜๋‹ค. ํ…Œ์ŠคํŠธ์—๋Š” 3๊ฐ€์ง€ ๋ฐฉ๋ฒ• ๋ชจ๋‘ ๊ฐ™์€ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์˜€๊ณ , VOV+CNN ๋ฐฉ๋ฒ•์€ ์ „์ฒด ์ด๋ฏธ์ง€์— ๋Œ€ํ•ด์„œ ์œˆ๋„์šฐ๋ฅผ ์ด๋™ํ•˜๋ฉด์„œ ๊ฒ€์ถœ์„ ์‹œ๋„ํ•˜์˜€๋‹ค.

ํ‘œ 1. ํ•™์Šต๊ณผ ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ

Table 1. Training and test data

No. of Training data

(VOV+CNN)

No. of Training data

(Faster R-CNN, YOLOv2)

No. of Test data

Defects

56,000

2,600

800

Non-

defects

56,000

-

2,300

VOV์™€ CNN์„ ๊ฒฐํ•ฉํ•œ ๊ฒ€์ถœ ๋ฐฉ๋ฒ•์€ ๊ธˆ์† ๋ถ€ํ’ˆ์˜ ํฐ ์ด๋ฏธ์ง€๋ฅผ ์Šฌ๋ผ์ด๋”ฉ ์œˆ๋„์šฐ ๋ฐฉ์‹์œผ๋กœ ์ž˜๋ผ์„œ ์ž‘์€ ํฌ๊ธฐ์˜ ROI ์ด๋ฏธ์ง€๋ฅผ ์ƒ์„ฑํ•œ๋‹ค. ์ƒ์„ฑ๋œ ROI ์ด๋ฏธ์ง€์— VOV ํ•„ํ„ฐ๋ฅผ ์ ์šฉํ•ด ๊ฒฐํ•จ ํ›„๋ณด์˜์—ญ์„ ์ถ”์ถœํ•˜๊ณ , ์ด๋ฅผ CNN์— ์ž…๋ ฅ์œผ๋กœ ์ฃผ์–ด ํ•™์Šตํ•œ๋‹ค. Faster R-CNN ๋ฐ YOLOv2์—์„œ์˜ ๊ฒ€์ถœ์€์œ„ ๋ฐฉ๋ฒ•๊ณผ๋Š” ๋‹ฌ๋ฆฌ ROI ์ด๋ฏธ์ง€ ๋Œ€์‹ ์— ํฐ ์ด๋ฏธ์ง€๋ฅผ ํ•™์Šต์— ์‚ฌ์šฉํ•œ๋‹ค. ํฐ ์ด๋ฏธ์ง€์—์„œ ๊ฒฐํ•จ์ด ์žˆ๋Š” ์ขŒํ‘œ๋ฅผ ๋ฐ์ดํ„ฐํ™” ํ•œ ํ›„, ํฐ ์ด๋ฏธ์ง€์™€ ๊ฒฐํ•จ ์ขŒํ‘œ ๋ฐ์ดํ„ฐ๋ฅผ ์ž…๋ ฅ์œผ๋กœ ์‚ฌ์šฉํ•ด ํ•™์Šตํ•œ๋‹ค. ํ‰๊ฐ€์‹œ, Faster R-CNN์€ ๊ฒฐํ•จ์˜์—ญ์˜ ์ขŒ์ธก ์ƒ๋‹จ ์ขŒํ‘œ, YOLOv2๋Š” ๊ฒฐํ•จ์˜์—ญ์˜ ์ค‘์‹ฌ ์ขŒํ‘œ๋ฅผ ์ถœ๋ ฅํ•œ๋‹ค. YOLOv2๋Š” ์ƒ๋Œ€์ ์ธ ์ขŒํ‘œ(์ขŒํ‘œ/์ „์ฒด ํ•ด์ƒ๋„)๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค.

VOV+CNN ๊ธฐ๋ฐ˜ ๊ฒฐํ•จ ๊ฒ€์ถœ ์‹คํ—˜์€ Epoch 300, ํ•™์Šต๋ฅ  10-5 ์ˆ˜ํ–‰๋˜์—ˆ์œผ๋ฉฐ, Faster R-CNN ํ•™์Šต์€ 4 ์„ธ๋ถ€ ๋‹จ๊ณ„๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ ๊ฐ ๋‹จ๊ณ„๋‹น 100์„ธ๋Œ€(epoch)์”ฉ, ํ•™์Šต๋ฅ  10-5๋กœ ์ˆ˜ํ–‰๋˜์—ˆ๋‹ค. YOLOv2 ํ•™์Šต์€ ์„ธ๋ถ€๋‹จ๊ณ„ ์—†์ด ์ „์ฒด 300์„ธ๋Œ€ ๋งŒํผ ์ˆ˜ํ–‰ํ•˜๋ฉฐ, ํ•™์Šต๋ฅ ์€ 0.001๋ถ€ํ„ฐ ์‹œ์ž‘ํ•˜์—ฌ 10-5๊นŒ์ง€ ๋‹จ๊ณ„์ ์œผ๋กœ ๊ฐ์†Œํ•œ๋‹ค. ์‹คํ—˜์— ์‚ฌ์šฉ๋œ ๋„คํŠธ์›Œํฌ์˜ ๊ตฌ์กฐ๋Š” ๊ทธ๋ฆผ 5์— ๋‚˜์™€ ์žˆ๋‹ค.

YOLOv2 ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ์ ์šฉ ์‹œ ์•ต์ปค ๋ฐ•์Šค์˜ ์ˆ˜๋Š” ์ „์ฒด ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ k-means ํด๋Ÿฌ์Šคํ„ฐ๋ง์„ ํ†ตํ•ด ๊ฒฐ์ •ํ•œ๋‹ค. ์ฐธ๊ณ ๋ฌธํ—Œ(14)๊ณผ ๋™์ผํ•œ ๋ฐฉ๋ฒ•์œผ๋กœ ๊ฒฐํ•จ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ์ˆ˜ํ–‰ํ•œ ๊ฒฐ๊ณผ k=5์—์„œ ํ‰๊ท  IOU(Intersection over Union)๊ฐ€ 0.61๋กœ์„œ k๊ฐ€ ๋” ์ฆ๊ฐ€๋˜์–ด๋„ IOU์˜ ํ–ฅ์ƒ์ด ๋‘”ํ™”๋˜๋ฏ€๋กœ, ์•Œ๊ณ ๋ฆฌ์ฆ˜ ์ˆ˜ํ–‰์˜ ๋ณต์žก๋„๋„ ์ค„์ผ ๊ฒธ, 5๊ฐœ์˜ ์•ต์ปค ๋ฐ•์Šค๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค.

์‹คํ—˜์—์„œ ๊ณ„์‚ฐํ•˜๋Š” ์„ฑ๋Šฅ ์ง€ํ‘œ๋Š” ํฌ๊ฒŒ ๋‘ ์ข…๋ฅ˜์ด๋‹ค. ์ฒซ์งธ๋Š” ๋ฏธ๊ฒ€์ถœ์œจ(FNR), ๊ณผ๊ฒ€์ถœ์œจ(FPR), ์˜ค๋ฅ˜์œจ(Error Rate)๋กœ์„œ ์‚ฐ์—…์ฒด ํ˜„์žฅ์—์„œ ๋งŽ์ด ์“ฐ์ด๋ฉฐ ์ง๊ด€์  ์ดํ•ด๊ฐ€ ์‰ฝ๋‹ค. ๋‘˜์งธ๋Š” ์žฌํ˜„์œจ(Recall), ์ •๋ฐ€๋„(Precision), ์ •ํ™•๋„(Accuracy)๋กœ์„œ ์—ฐ๊ตฌ ๋…ผ๋ฌธ์—์„œ ๋งŽ์ด ์“ฐ์ด๋Š” ์ง€ํ‘œ๋“ค์ด๋‹ค. ํ‘œ 2์— ์ด๋“ค ์„ฑ๋Šฅ ์ง€ํ‘œ์— ๋Œ€ํ•œ ๋น„๊ต๊ฐ€ ์ •๋ฆฌ๋˜์–ด ์žˆ๋‹ค. ๊ธˆ์† ๋ถ€ํ’ˆ์˜ ํ‘œ๋ฉด ๊ฒฐํ•จ์— ๋Œ€ํ•œ ๊ฒ€์ถœ ์‹คํ—˜ ๊ฒฐ๊ณผ๊ฐ€ ํ‘œ 3๊ณผ ๊ทธ๋ฆผ 7์— ๋‚˜์™€ ์žˆ๋‹ค. ํ‘œ 3์—์„œ Faster R-CNN ๊ธฐ๋ฐ˜ ๊ฒฐํ•จ ๊ฒ€์ถœ ๋ฐฉ๋ฒ•์ด VOV+CNN ๋ฐฉ๋ฒ•์— ๋น„ํ•ด ๋ฏธ๊ฒ€์ถœ์œจ์ด(FNR) ์•ฝ 1.1% ๊ฐ์†Œ, ๊ณผ๊ฒ€์ถœ์œจ์ด(FPR) ์•ฝ 56.1% ๊ฐ์†Œ, ์˜ค๋ฅ˜์œจ์ด(Error Rate) ์•ฝ 42.2% ๊ฐ์†Œํ•˜์—ฌ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์˜€๋‹ค. ๊ฒ€์‚ฌ์‹œ๊ฐ„ ์ธก๋ฉด์—์„œ๋„ Faster R-CNN์ด ๋” ์ ๊ฒŒ ๊ฑธ๋ฆผ์„ ์•Œ ์ˆ˜ ์žˆ๋‹ค. YOLOv2 ๋ฐฉ๋ฒ•์ด Faster R-CNN ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ•๋ณด๋‹ค ๋ฏธ๊ฒ€์ถœ์œจ์€ 3.3% ์ฆ๊ฐ€ํ–ˆ์œผ๋‚˜, ๊ณผ๊ฒ€์ถœ์œจ๊ณผ ์˜ค๋ฅ˜์œจ์€ ๊ฐ๊ฐ 13.2% ๋ฐ 8.3% ๊ฐ์†Œํ•˜์˜€๋‹ค. ํŠนํžˆ, ๊ฒ€์‚ฌ ์‹œ๊ฐ„๊ณผ ์ดˆ๋‹น ์ฒ˜๋ฆฌ ํ”„๋ ˆ์ž„ ์ˆ˜๋Š” 3๊ฐ€์ง€ ๋ฐฉ๋ฒ• ์ค‘ ํ˜„๊ฒฉํ•œ ์ฐจ์ด๋กœ ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•˜์˜€๋‹ค. ๊ฒ€์‚ฌ์‹œ๊ฐ„ ์ธก์ •์— ์‚ฌ์šฉ๋œ ๊ทธ๋ž˜ํ”ฝ์นด๋“œ๋Š” GTX 960์ด๋‹ค. ์ฝ”๋“œ ๊ตฌํ˜„์€ VOV+CNN, Faster R-CNN๋Š” Matlab์„ ์‚ฌ์šฉํ•˜์˜€๊ณ , YOLOv2๋Š” darknet19 ๊ธฐ๋ฐ˜์˜ C/C++๋ฅผ ์‚ฌ์šฉํ•˜์˜€๋‹ค. Faster R-CNN๊ณผ YOLOv2 ์ˆ˜ํ–‰ ์†๋„์˜ ์ฐจ์ด๊ฐ€ ๊ธฐ์กด์˜ ์ฐธ๊ณ ๋ฌธํ—Œ(14)์˜ ๊ฒฐ๊ณผ์™€ ์œ ์‚ฌํ•˜๊ฒŒ ๋‚˜์™”๋‹ค.

ํ‘œ 2. ์„ฑ๋Šฅ ์ง€ํ‘œ ์ •๋ฆฌ

Table 2. Summary of perfomrnce indexes

๋ฏธ๊ฒ€์ถœ์œจ

(FNR)

F N P = F N T P + F N

์žฌํ˜„์œจ

(Recall)

T P P = T P T P + F N

๊ณผ๊ฒ€์ถœ์œจ

(FPR)

F P N = F P T N + F P

์ •๋ฐ€๋„

(Precision)

T P T P + F P

์˜ค๋ฅ˜์œจ

(Error Rate)

1 - T P + T N P + N

์ •ํ™•๋„

(Accuracy)

T P + T N P + N

ํ‘œ 3. ๊ธˆ์† ๋ถ€ํ’ˆ ํ‘œ๋ฉด์˜ ๊ฒฐํ•จ ๊ฒ€์ถœ ์‹คํ—˜ ๊ฒฐ๊ณผ 1

Table 3. Experimental results 1 of defect detection for metal parts

VOV+CNN

Faster R-CNN

YOLO V2

False Negative Rate (FNR)

2 %

0.9 %

4.2 %

False Positive

Rate (FPR)

79.1 %

23 %

9.8 %

Error Rate

59.6 %

17.4 %

9.1 %

Inspection time (100 images)

18.9 sec

16.8 sec

3.4 sec

fps(frame per second)

5.2

5.9

29.4

๊ทธ๋ฆผ. 6. ๋„คํŠธ์›Œํฌ ๊ตฌ์กฐโ€“์œ„) VOV+CNN, Faster R-CNN, ์•„๋ž˜) YOLOv2

Fig. 6. Network structureโ€“top) VOV+CNN, Faster R-CNN, bottom) YOLOv2

../../Resources/kiee/KIEE.2018.67.7.865/fig6.png

๊ทธ๋ฆผ 7์€ 3๊ฐ€์ง€ ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์žฌํ˜„์œจ(Recall), ์ •๋ฐ€๋„(Precision), ์ •ํ™•๋„(Accuracy) ์„ฑ๋Šฅ์„ ๊ทธ๋ž˜ํ”„๋กœ ๋‚˜ํƒ€๋‚ธ ๊ฒƒ์ด๋‹ค. ์ •ํ™•๋„์™€ ์ •๋ฐ€๋„๋Š” YOLOv2 ๋ฐฉ๋ฒ•์ด ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•˜๊ณ , ์žฌํ˜„์œจ์—์„œ๋Š” Faster R-CNN ๋ฐฉ๋ฒ•์ด ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•˜๊ฒŒ ๋‚˜์™”๋‹ค.

๊ทธ๋ฆผ 8์—๋Š” Faster R-CNN๊ณผ YOLOv2์˜ ๊ฒฐํ•จ ๊ฒ€์ถœ ์˜ˆ๊ฐ€ ๋‚˜์™€ ์žˆ๋‹ค. ๋‘ ๋ฐฉ๋ฒ• ๊ฐ„์— ๊ฒฐํ•จ ํ‘œ์‹œ ์‚ฌ๊ฐํ˜•์˜ ๊ฐœ์ˆ˜ ๋ฐ ํ˜•ํƒœ์˜ ์ฐจ์ด๊ฐ€ ๋‚˜ํƒ€๋‚จ์„ ์•Œ ์ˆ˜ ์žˆ๋‹ค. Faster R-CNN ๊ฒ€์ถœ์€ ๊ฐ™์€ ๊ฒฐํ•จ์— ๋Œ€ํ•ด 9๊ฐœ์˜ ์•ต์ปค ์ค‘ ์ผ๋ถ€๊ฐ€ ์‚ฌ์šฉ๋˜์–ด ์ค‘๋ณต๋œ ๊ฒ€์ถœ์ด ์‚ฌ๊ฐํ˜•์œผ๋กœ ํ‘œ์‹œ๋˜๋Š”๋ฐ, ๊ฐ™์€ ๊ฒฐํ•จ์„ ํฌํ•จํ•˜๊ณ  ์žˆ๋‹ค๋ฉด ํ•˜๋‚˜์˜ ์˜์—ญ์œผ๋กœ ๊ฐ„์ฃผํ•˜์—ฌ ์‹คํ—˜ ๊ฒฐ๊ณผ์— ๋ฐ˜์˜ํ•˜์˜€๋‹ค. YOLOv2์—์„œ๋Š” ๊ฒฐํ•จ ๊ฒ€์ถœ ์ˆ˜๊ฐ€ ์ ๊ฒŒ ๋‚˜์™”๊ณ , ์ด๋ฅผ ํ†ตํ•ด ๊ณผ๊ฒ€์ถœ ์ง€ํ‘œ๊ฐ€ ๊ฐ์†Œํ•˜์˜€์Œ์„ ์œ ์ถ”ํ•  ์ˆ˜ ์žˆ๋‹ค. ๋Œ€์‹ , ์šฐ์ธก ์ƒ๋‹จ์—์„œ ๊ธธ์ด๊ฐ€ ๊ธด ์‚ฌ๊ฐํ˜•์œผ๋กœ ๊ฒ€์ถœ์ด ํ‘œ์‹œ๋˜์—ˆ๋Š”๋ฐ, ๊ฒฐํ•จ์„ ํฌํ•จํ•ด ์„ธ๋กœ๋กœ ๋‚˜ํƒ€๋‚˜ ์žˆ๋Š” ์„ ์€ ๊ฒฐํ•จ์ด ์•„๋‹ˆ์ง€๋งŒ ์Šคํฌ๋ž˜์น˜์™€ ๊ฐ™์€ ๊ฒฐํ•จ์œผ๋กœ ์ธ์‹๋˜์–ด ์‹ค์ œ๋ณด๋‹ค ๋„“์€ ์˜์—ญ์œผ๋กœ ํ‘œ์‹œ๋œ ๊ฒƒ์ด๋‹ค.

๊ทธ๋ฆผ. 7. ๊ฒฐํ•จ ๋ฐ์ดํ„ฐ์˜ k-means clustering ๊ฒฐ๊ณผ

Fig. 7. K-means clustering results for defect data

../../Resources/kiee/KIEE.2018.67.7.865/fig7.png

๊ทธ๋ฆผ. 8. ๊ธˆ์† ๋ถ€ํ’ˆ ํ‘œ๋ฉด์˜ ๊ฒฐํ•จ ๊ฒ€์ถœ ์‹คํ—˜ ๊ฒฐ๊ณผ 2

Fig. 8. Experimental results 2 of defect detection for metal parts

../../Resources/kiee/KIEE.2018.67.7.865/fig8.png

๊ทธ๋ฆผ. 9. ๊ฒฐํ•จ ๊ฒ€์ถœ ์˜ˆ - Faster R-CNN, YOLOv2

Fig. 9. Detection results of defects for for metal parts - Faster R-CNN and YOLOv2

../../Resources/kiee/KIEE.2018.67.7.865/fig9.png

5. ๊ฒฐ ๋ก 

์‚ฐ์—…์šฉ ๊ธˆ์† ๋ถ€ํ’ˆ์˜ ํ‘œ๋ฉด ๊ฒฐํ•จ ๊ฒ€์‚ฌ์— CNN ๊ธฐ๋ฐ˜์˜ 3๊ฐ€์ง€ ๊ฒ€์ถœ ๊ธฐ๋ฒ•์„ ์ ์šฉํ•˜์—ฌ ์‹คํ—˜ํ•˜์˜€๋‹ค. ์‚ฌ์ „์— ์—ฐ๊ตฌ๋œ VOV+CNN ๊ฒ€์ถœ ๋ฐฉ๋ฒ• ์™ธ์— Faster R-CNN๊ณผ YOLOv2 ๊ธฐ๋ฒ•์„ ์ค‘์‹ฌ์œผ๋กœ ๊ฒ€์ถœ ์„ฑ๋Šฅ๊ณผ ์†๋„ ์ธก๋ฉด์—์„œ ๋น„๊ตํ•˜์˜€๋‹ค. ๋ฏธ๊ฒ€์ถœ์œจ์€ Faster R-CNN์ด ๊ฐ€์žฅ ์šฐ์ˆ˜ํ–ˆ๊ณ , ๊ทธ ๋‹ค์Œ์œผ๋กœ VOV+CNN, YOLOv2 ์ˆœ์ด์—ˆ๋‹ค. ๊ณผ๊ฒ€์ถœ๊ณผ ์˜ค๋ฅ˜์œจ์€ YOLOv2๊ฐ€ ๊ฐ€์žฅ ์šฐ์ˆ˜ํ–ˆ๊ณ , ๊ทธ ๋‹ค์Œ์œผ๋กœ Faster R-CNN, VOV+CNN ์ˆœ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ๋‹ค. ๋˜ํ•œ, ๊ฒ€์ธจ์†๋„๋ฉด์—์„œ๋„ YOLOv2 ๊ธฐ๋ฒ•์ด ๋‹ค๋ฅธ ๋‘ ๋ฐฉ๋ฒ•๋ณด๋‹ค ์›”๋“ฑํ•œ ์„ฑ๋Šฅ์„ ๋ณด์˜€๋‹ค. ๋ฌผ์ฒด ๊ฒ€์ถœ์„ ์œ„์ฃผ๋กœ ์—ฐ๊ตฌ๊ฐ€ ๋˜๊ณ  ์žˆ๋Š” Faster R-CNN๊ณผ YOLOv2 ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๋‚œ์ด๋„๊ฐ€ ๋†’์€ ๊ธˆ์† ๋ถ€ํ’ˆ์˜ ํ‘œ๋ฉด ๊ฒฐํ•จ ๊ฒ€์ถœ์— ์ ์šฉํ•˜์—ฌ ๋งŒ์กฑํ•  ๋งŒํ•œ ์„ฑ๋Šฅ์„ ์–ป์€ ๊ฒƒ์ด ์˜์˜๋ผ๊ณ  ๋ณผ ์ˆ˜ ์žˆ๋‹ค.

ํ–ฅํ›„, ๋„คํŠธ์›Œํฌ ๊ตฌ์กฐ์˜ ์ตœ์ ํ™”์™€ ํ•™์Šต ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ๊ฐœ์„ ์„ ํ†ตํ•ด ๊ฒ€์ถœ ์„ฑ๋Šฅ๊ณผ ์ฒ˜๋ฆฌ ์‹œ๊ฐ„์— ๋Œ€ํ•œ ์ถ”๊ฐ€์ ์ธ ํ–ฅ์ƒ์ด ํ•„์š”ํ•˜๋‹ค.

References

1 
Park Y. , Kweon I. S. , 2016, Ambiguous Surface Defect Image Classification of AMOLED Displays in Smart- phones,, IEEE Trans. on Industrial Informatics, Vol. 12, No. 2, pp. 597-607DOI
2 
Ghorai S. , Mukherjee A. , Gangadaran M. , Dutta P. K. , 2013, Automatic Defect Detection on Hot-Rolled Flat Steel Products, IEEE Trans. on Instrumentation and Measurement, Vol. 62, No. 3, pp. 612-621DOI
3 
LeCun Yann, 1998, Gradient based learning applied to document recognition, Proceedings of the IEEE, pp. 2278-2324DOI
4 
Park J-K. , Kwon N. , J-H. , Kang D. , 2016, Machine Learning-Based Imaging System for Surface Defect Inspection, International Journal of Precision Engineering and Manufacturing-Green Technology, Vol. 3, No. 3, pp. 303-310DOI
5 
Choi H. , Seo K. , 2017, CNN Based Detection of Surface Defects for Electronic Parts, Journal of Korean Institute of Intelligent Systems, Vol. 27, No. 3, pp. 195-200Google Search
6 
Choi H. , Seo K. , 2017, Comparison of CNN Structures for Detection of Surface Defects, The Transactions of the Korean Institute of Electrical Engineers, Vol. 66, No. 7, pp. 1100-1104DOI
7 
Kwon B.-K , Won J.-S. , Kang D.-J. , 2015, Fast defect detection for various types of surfaces using random forest with VOV features, International Journal of Precision Engineering and Manufacturing, Vol. 16, No. 5, pp. 965-970DOI
8 
Girshick R. , Donahue J. , Darrell T. , Malik J. , 2016, Region- Based Convolutional Networks for Accurate Object Detection and Segmentation, IEEE Trans. on Pattern Analysis and Machine Intelligence, Vol. 38, No. 1, pp. 142-158DOI
9 
Girshick R. , , Fast R-CNN, ICCV 2015, pp. 1440-1448DOI
10 
Ren Shaoqing, He Kaiming , Girshick Ross , Sun Jian , 2015, Faster r-cnn: Towards real-time object detection with region proposal networks, Advances in neural information processing systems, pp. 91-99DOI
11 
Lee M. , Seo K, 2017, Using Deep Learning Based Machine Vision for Defects Detection,, Proceedings of Information and Control Symposium CICS'2017, pp. 54-55Google Search
12 
Redmon Joseph , Divvala Santosh , Girshick Ross , Farhadi Ali , 2016, You Only Look Once: Unified, Real-Time Object Detection, Proceedings of the IEEE conference on computer vision and pattern recognitionGoogle Search
13 
Lee M. , Seo K, 2018, Comparison of CNN Methods for Defects Detection, Proceedings of Information and Control Symposium ICS'2018, pp. 101-102Google Search
14 
Redmon Joseph , Farhadi Ali , 2017, YOLO9000: better, faster, stronger, arXiv:1612.08242Google Search

์ €์ž์†Œ๊ฐœ

์ด ๋ฏผ ๊ธฐ (Minki Lee)
../../Resources/kiee/KIEE.2018.67.7.865/au1.png

2017๋…„ ์„œ๊ฒฝ๋Œ€ํ•™๊ต ์ „์ž๊ณตํ•™๊ณผ ์กธ์—…(ํ•™์‚ฌ)

2017๋…„~ํ˜„์žฌ ์„œ๊ฒฝ๋Œ€ํ•™๊ต ๋Œ€ํ•™์› ์ „์ž์ปดํ“จํ„ฐ๊ณตํ•™๊ณผ ์„์‚ฌ๊ณผ์ •

๊ด€์‹ฌ๋ถ„์•ผ:๋”ฅ๋Ÿฌ๋‹, ์˜์ƒ์ธ์‹, ๋จธ์‹ ๋น„์ „

์„œ ๊ธฐ ์„ฑ (Kisung Seo)
../../Resources/kiee/KIEE.2018.67.7.865/au2.png

1993๋…„ ์—ฐ์„ธ๋Œ€ํ•™๊ต ๋Œ€ํ•™์› ์ „๊ธฐ๊ณตํ•™๊ณผ ์กธ์—…(๋ฐ•์‚ฌ)

1999~2003๋…„ Michigan State University, Genetic Algorithms Research and Applications Group, Research Associate

2002~2003๋…„ Michigan State University, Electrical & Computer Engineering, Visiting Assistant Professor

2011~2012๋…„ Michigan State University, BEACON (Bio/ computational Evolution in Action CONsortium) Center, Visiting Scholar

1993๋…„~ํ˜„์žฌ ์„œ๊ฒฝ๋Œ€ํ•™๊ต ์ „์ž๊ณตํ•™๊ณผ ๊ต์ˆ˜

๊ด€์‹ฌ๋ถ„์•ผ: ์ง„ํ™”์—ฐ์‚ฐ, ๋”ฅ๋Ÿฌ๋‹, ์˜์ƒ์ธ์‹, ๋จธ์‹ ๋น„์ „, ๊ธฐ์ƒ์˜ˆ์ธก, ์ง€๋Šฅ๋กœ๋ด‡