• ๋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 Electrical Engineering, Kwangwoon University, Korea)
  2. (Department of Electronics Engineering, Seokyeong University, Korea)



Defects detection, Convolutional neural network, Network Reduction, Embedded System, YOLOv2, YOLOv3, YOLOv2-tiny

1. ์„œ ๋ก 

์˜์ƒ ๊ธฐ๋ฐ˜์˜ ๊ฒฐํ•จ ๊ฒ€์‚ฌ ๊ธฐ์ˆ ์ด ๋ฐœ๋‹ฌํ•˜์—ฌ ๋‹ค์–‘ํ•œ ๋ถ„์•ผ์—์„œ ๊ฒ€์‚ฌ ์ž๋™ํ™”์˜ ์ ์šฉ์ด ๋น ๋ฅด๊ฒŒ ์ฆ๊ฐ€ํ•˜๊ณ  ์žˆ๋‹ค. ํŠนํžˆ, ๊ธฐ์กด์˜ ๋จธ์‹ ๋น„์ „ ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ์„ฑ๋Šฅ์ด ๋„๋‹ฌํ•˜์ง€ ๋ชปํ•œ ์ˆ˜์ค€์„ ์„ ๋ณด์ด๋ฉด์„œ ๋‚œ์ด๋„๊ฐ€ ๋†’์€ ๋Œ€์ƒ์— ๋Œ€ํ•ด์„œ๋„ ์ ์šฉ ๋ฒ”์œ„๋ฅผ ๋Š˜๋ ค๊ฐ€๊ณ  ์žˆ๋‹ค[1-3].

ํ‘œ๋ฉด์˜ ๊ฒฐํ•จ ๊ฒ€์‚ฌ ๋ฐฉ์‹์€ ๊ฒ€์‚ฌ ํ™˜๊ฒฝ์— ๋”ฐ๋ผ ํฌ๊ฒŒ 3๊ฐ€์ง€๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์žˆ๋‹ค. ์ฒซ์งธ, ๊ณ ์ •๋œ ์œ„์น˜์—์„œ ๋ถ€ํ’ˆ์„ ๊ฒ€์‚ฌํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ์‚ฐ์—…์šฉ ๋จธ์‹ ๋น„์ „ ์‹œ์Šคํ…œ์—์„œ ์ฃผ๋กœ ์‚ฌ์šฉ๋˜๊ณ  ์žˆ๋‹ค. ๋‘˜์งธ, ๋กœ๋ด‡์— ๋ถ€์ฐฉ๋œ ๊ฒ€์‚ฌ์žฅ์น˜๋ฅผ ์ด๋™ํ•˜์—ฌ ์ œํ•œ๋œ ์ž‘์—… ์˜์—ญ๋‚ด์—์„œ์˜ ํ™•์žฅ๋œ ์œ„์น˜๋“ค์„ ๊ฒ€์‚ฌํ•˜๋Š” ๊ฒƒ์œผ๋กœ, ๋ถ€ํ’ˆ์˜ ์ „ํ›„์ขŒ์šฐ ๋“ฑ์˜ ์—ฌ๋Ÿฌ ๊ณณ์„ ๊ฒ€์‚ฌํ•˜๋Š”๋ฐ ์‚ฌ์šฉ๋œ๋‹ค. ์…‹์งธ, ๋“œ๋ก  ๋“ฑ์˜ ์ด๋™์ฒด๊ฐ€ ์ด๋™ํ•˜๋ฉด์„œ ๋„“์€ ์˜์—ญ์— ๋ถ„ํฌ๋˜์–ด์žˆ๋Š” ๊ฒฐํ•จ๋“ค์„ ๊ฒ€์‚ฌํ•œ๋‹ค. ๊ฑด๋ฌผ ์™ธ๋ฒฝ, ๋น„ํ–‰๊ธฐ์™€ ์„ ๋ฐ• ๋ณธ์ฒด, ํฐ ์ฒ ํŒ ํ‘œ๋ฉด ๋“ฑ์˜ ๊ฒฐํ•จ์„ ๊ฒ€์‚ฌํ•˜๋Š” ์˜ˆ๋ฅผ ๋“ค ์ˆ˜ ์žˆ๋‹ค.

๊ทธ๋ฆผ 1์— 3๊ฐ€์ง€ ํ‘œ๋ฉด ๊ฒฐํ•จ ๊ฒ€์‚ฌ ๋ฐฉ์‹์— ๋Œ€ํ•œ ์˜ˆ์‹œ๋„๊ฐ€ ๋‚˜์™€ ์žˆ๋‹ค. ๊ณ ์ •ํ˜• ๋ฐฉ์‹์ด ์ฃผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๊ธˆ์† ์ผ€์ด์Šค, LED ํŒจ๋„, ์ „์ž์†Œ์ž์˜ ํ‘œ๋ฉด ๊ฒฐํ•จ์— ๋Œ€ํ•ด์„œ๋Š” ์ƒ์‚ฐ ๊ทœ๋ชจ๊ฐ€ ํฌ๊ธฐ ๋•Œ๋ฌธ์— ์ด๋ฏธ ์ƒ๋‹น ๋ถ€๋ถ„ ๊ฒ€์‚ฌ ์ž๋™ํ™”๊ฐ€ ์ด๋ฃจ์–ด์ง€๊ณ  ์žˆ๋‹ค. ํ•œ ์ œํ’ˆ์— ๋Œ€ํ•ด์„œ ๋‹ค์ˆ˜์˜ ์œ„์น˜๋ฅผ ์ด๋™ํ•˜๋ฉฐ ๊ฒ€์‚ฌํ•ด์•ผ ํ•˜๋Š” ๊ฒฝ์šฐ ๋กœ๋ด‡ํ˜•์ด ์“ฐ์ด๋ฉฐ, ๋กœ๋ด‡ํŒ”์ด ๊ฒ€์‚ฌ ๋Œ€์ƒ ๋ถ€์œ„๋ฅผ ์ด๋™ํ•˜๋ฉด์„œ ๊ฒ€์‚ฌํ•œ๋‹ค. ์ด ๋ฐฉ์‹๋„ ์ž๋™์ฐจ ๋ฐ ์ „์ž ๋ถ„์•ผ์—์„œ ์ ์šฉ์ด ํ™•๋Œ€๋˜๊ณ  ์žˆ๋‹ค.

๊ทธ๋ฆผ. 1. ํ‘œ๋ฉด ๊ฒฐํ•จ ๊ฒ€์‚ฌ ๋ฐฉ์‹์˜ ์ข…๋ฅ˜ a) ๊ณ ์ •, b) ๋กœ๋ด‡, c) ๋“œ๋ก 

Fig. 1. Types of surface defects inspection a) fixed, b) robot, c) drone

../../Resources/kiee/KIEE.2020.69.2.325/fig1.png

๊ทธ๋Ÿฌ๋‚˜ ๊ฒ€์‚ฌ ์˜์—ญ์˜ ๊ทœ๋ชจ๊ฐ€ ํฐ ๊ฑด๋ฌผ ์™ธ๋ฒฝ์ด๋‚˜ ๋Œ€ํ˜• ๊ตฌ์กฐ๋ฌผ์˜ ์ฒ ํŒ ํ‘œ๋ฉด ๋“ฑ์— ๋Œ€ํ•œ ๊ฒ€์‚ฌ๋Š” ์•„์ง๊นŒ์ง€ ์ž๋™ํ™”์˜ ์ •๋„๊ฐ€ ๋ฏธํกํ•œ ํŽธ์ด๋‹ค. ๋„“์€ ๊ฒ€์‚ฌ ์˜์—ญ์„ ์ด๋™ํ•˜๋Š” ๋“œ๋ก  ๋“ฑ์˜ ์ด๋™์ฒด๊ฐ€ ํ•„์š”ํ•˜๊ณ  ๊ฒ€์‚ฌ ํ”Œ๋žซํผ๋„ ๋ฐ์Šคํฌํƒ‘ ๊ธฐ๋ฐ˜์˜ ๊ฒ€์‚ฌ์žฅ์น˜๊ฐ€ ์•„๋‹ˆ๋ผ ์ž„๋ฒ ๋””๋“œํ˜• ์žฅ์น˜๋ฅผ ํ•„์š”๋กœ ํ•œ๋‹ค. ๋˜ํ•œ ๊ฐ€๊ณต ๋งˆ๋ฌด๋ฆฌ๊ฐ€ ์†Œ๊ทœ๋ชจ ๋ถ€ํ’ˆ์— ๋น„ํ•ด ๊ฑฐ์น ๊ณ , ์‚ฌ์šฉ์ƒ ๋ฌธ์ œ๊ฐ€ ์—†์œผ๋ฉด ํ‘œ๋ฉด์˜ ๋ถˆ์ˆœ๋ฌผ์ด๋‚˜ ์˜ค์—ผ๋“ฑ์ด ํ˜ผ์žฌํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ๋‹ค. ๋˜ํ•œ ๊ฒฐํ•จ๊ณผ ๋น„๊ฒฐํ•จ์˜ ๊ตฌ๋ณ„์ด ๋งค์šฐ ์–ด๋ ค์›Œ์„œ ๊ฒ€์ถœ ๋‚œ์ด๋„๋„ ๋†’๋‹ค. ๋”ฐ๋ผ์„œ ์ „ํ†ต์ ์ธ ๋จธ์‹ ๋น„์ „ ๊ธฐ๋ฒ•์˜ ์„ฑ๋Šฅ ํ•œ๊ณ„๋ฅผ ๋„˜์„ ์ˆ˜ ์žˆ๋Š” ๋”ฅ๋Ÿฌ๋‹ ๊ธฐ๋ฒ•์˜ ์ ์šฉ์ด ์š”๊ตฌ๋˜๋ฉฐ, ๋™์‹œ์— ์ž„๋ฒ ๋””๋“œ ์‹œ์Šคํ…œ์—์„œ ์ฒ˜๋ฆฌ๊ฐ€ ๊ฐ€๋Šฅํ•ด์•ผ ํ•˜๋Š” ์ œํ•œ ์‚ฌํ•ญ์ด ์žˆ๋‹ค.

๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ๋“œ๋ก  ๋“ฑ์˜ ์ด๋™์ฒด๋ฅผ ํ†ตํ•ด ๊ฒ€์‚ฌ๋ฅผ ์ง„ํ–‰ํ•˜๋Š” ํฐ ์ฒ ํŒ ๋˜๋Š” ๊ธˆ์†ํŒ์˜ ํ‘œ๋ฉด ๊ฒ€์‚ฌ๋ฅผ ๋Œ€์ƒ์œผ๋กœ ํ•˜์—ฌ, ์ด์— ํ•„์š”ํ•œ ์š”๊ตฌ ์กฐ๊ฑด๊ณผ ๋ฌธ์ œ ํŠน์„ฑ, ๊ทธ๋ฆฌ๊ณ  ๋”ฅ๋Ÿฌ๋‹ ๊ธฐ๋ฒ•๊ณผ ์ด์˜ ์ž„๋ฒ ๋””๋“œ ์‹œ์Šคํ…œ์—์˜ ์ ์šฉ์„ ๊ฒ€ํ† ํ•œ๋‹ค.

๋”ฅ๋Ÿฌ๋‹ ๊ธฐ๋ฐ˜ ๊ฒฐํ•จ ๊ฒ€์‚ฌ์— ๋Œ€ํ•œ ์„ ํ–‰ ์—ฐ๊ตฌ์—์„œ VOV[4] ํ•„ํ„ฐ์™€ CNN์„ ๊ฒฐํ•ฉํ•˜์—ฌ ํ‘œ๋ฉด ๊ฒฐํ•จ์— ๋Œ€ํ•œ ๊ฒ€์ถœ์„ ์‹œ๋„ํ•˜์˜€๋‹ค[5]. ์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[6] ๊ธฐ๋ฐ˜ ์ ‘๊ทผ๋ฒ• ์ค‘ ๊ฐ€์žฅ ์ตœ์‹  ๋ฐฉ๋ฒ•์ธ Faster R-CNN[7]์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฒฐํ•จ์˜ ๊ฒ€์ถœ ์„ฑ๋Šฅ๊ณผ ์†๋„๋ฅผ ํ–ฅ์ƒ์‹œ์ผฐ๋‹ค[8]. ๋˜ํ•œ, YOLOv2[9] ๊ธฐ๋ฒ•์„ ์ ์šฉํ•˜์—ฌ Faster R-CNN ์ ‘๊ทผ๋ฒ•๊ณผ ๋น„๊ตํ•œ ์—ฐ๊ตฌ๋„ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค[10]. ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ์‚ฌ์ „ ์—ฐ๊ตฌ๋ฅผ ํ™•์žฅํ•˜์—ฌ, ์ฒ ํŒ ํ‘œ๋ฉด์šฉ ๊ฒฐํ•จ ๊ฒ€์‚ฌ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด์„œ YOLOv2 ๋ฐ v3๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ž„๋ฒ ๋””๋“œ ์‹œ์Šคํ…œ ์ ์šฉ์„ ์œ„ํ•œ ๋„คํŠธ์›Œํฌ ์ถ•์†Œ ๋ชจ๋ธ์„ ๊ตฌํ˜„ํ•œ๋‹ค. NEU ์ฒ ํŒ ๋ฐ์ดํ„ฐ[11]๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ NVIDIA์‚ฌ์˜ Jetson TX1[12] ๋ณด๋“œ ํ™˜๊ฒฝ์—์„œ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•œ๋‹ค.

2. ๊ฒฐํ•จ ๋ฐ์ดํ„ฐ ํŠน์„ฑ ๋ฐ ์œ„์น˜ ๊ฒ€์ถœ

์ฒ ํŒ ๋“ฑ์˜ ๊ธˆ์† ํ‘œ๋ฉด ๊ฒฐํ•จ ๊ฒ€์‚ฌ๋Š” ํ‘œ๋ฉด์ด ๊ท ์ผํ•˜์ง€ ์•Š๊ฑฐ๋‚˜ ๋งค๋„๋Ÿฝ์ง€ ์•Š์€ ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์•„์„œ ํฌ๋ž™, ์Šคํฌ๋ž˜์น˜, ํ ์ง‘ ๋“ฑ์˜ ๊ฒฐํ•จ์ด ํ‘œ๋ฉด ๋ฐฐ๊ฒฝ๊ณผ ๊ตฌ๋ณ„ํ•˜๊ธฐ ๋งค์šฐ ์–ด๋ ต๋‹ค.

๊ทธ๋ฆผ. 2. ํ‘œ๋ฉด ๊ฒฐํ•จ์˜ ์˜ˆ

Fig. 2. Exmaples of surface defects

../../Resources/kiee/KIEE.2020.69.2.325/fig2.png

๊ทธ๋ฆผ 2๋Š” NEU ์ฒ ํŒ ๋ฐ์ดํ„ฐ์˜ ์ผ๋ถ€ ์˜ˆ์ด๋‹ค. ์žฌ์งˆ ํŠน์„ฑ๊ณผ ๊ฐ€๊ณต ๊ณผ์ •์—์„œ ํฌํ•จ๋˜๋Š” ๋ฐฐ๊ฒฝ ๋ฌด๋Šฌ๋“ค๊ณผ ์˜ค์—ผ ๋ฐ ๋ถˆ์ˆœ๋ฌผ๋กœ ์ธํ•ด์„œ ๊ฒฐํ•จ๊ณผ ์ •์ƒ์˜ ๊ตฌ๋ถ„์ด ์–ด๋ ต๊ณ  ์กฐ๋ช…์— ๋”ฐ๋ผ ๊ฒฐํ•จ์˜ ๊ฒ€์ถœ์ด ๋”์šฑ ์–ด๋ ค์šด ๊ฒฝ์šฐ๊ฐ€ ๋ฐœ์ƒํ•œ๋‹ค. ํ‘œ๋ฉด ๊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˜ ์ข…๋ฅ˜์™€ ์ƒ๊ด€์—†์ด ๊ฒฐํ•จ ๋ฐ ์œ„์น˜๋ฅผ ํŒ๋ณ„ํ•ด์•ผ ํ•˜๋Š” ๊ฒฐํ•จ ๊ฒ€์ถœ ๋ฌธ์ œ๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์žˆ๋‹ค. ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ํ›„์ž๋ฅผ ๋‹ค๋ฃฌ๋‹ค.

3. CNN ๊ธฐ๋ฐ˜ ๊ฒฐํ•จ ๊ฒ€์ถœ ๊ธฐ๋ฒ•

3.1 CNN VoV, Faster R-CNN

๋จธ์‹  ๋น„์ „ ๊ธฐ๋ฒ•์œผ๋กœ๋Š” ์ „ํ†ต์ ์ธ ํ•„ํ„ฐ ๊ธฐ๋ฐ˜์˜ ์ ‘๊ทผ๋ฒ•[1,2], ํ•„ํ„ฐ์™€ SVM๋“ฑ์˜ ๊ธฐ๊ณ„ํ•™์Šต์„ ๊ฒฐํ•ฉํ•œ ์ ‘๊ทผ๋ฒ•[2,3,4], ๊ทธ๋ฆฌ๊ณ  CNN(Convolutional Neural Network) ๋“ฑ์„ ์ด์šฉํ•œ ๋”ฅ๋Ÿฌ๋‹ ์ ‘๊ทผ๋ฒ•์ด ์žˆ๋‹ค[5,8]. ๋‚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.

3.2 YOLO v2, v3

YOLOv2[9]๋Š” ์‹ค์‹œ๊ฐ„ ๊ฐ์ฒด ๊ฒ€์ถœ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” ํ•™์Šต ๋ฐ ๊ฒ€์ถœ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด๋‹ค. ์ž…๋ ฅ ์ด๋ฏธ์ง€๋ฅผ ๊ทธ๋ฆฌ๋“œ ์…€๋กœ ๋‚˜๋ˆˆ ํ›„ ๊ฐ ์…€๋งˆ๋‹ค n๊ฐœ์˜ ์•ต์ปค ๋ฐ•์Šค๋ฅผ ์ ์šฉํ•˜์—ฌ ๊ทธ๋ฆฌ๋“œ ์…€ ๋‚ด์˜ ๊ฐ์ฒด์˜ ์กด์žฌ ํ™•๋ฅ , ๊ฐ์ฒด์— ๋Œ€ํ•œ ํด๋ž˜์Šค์˜ ํ™•๋ฅ ์„ ๊ตฌํ•œ๋‹ค. ์ด๋ฅผ ํ† ๋Œ€๋กœ ์ตœ์ข…์ ์œผ๋กœ ๊ฐ์ฒด์˜ ์ค‘์‹ฌ์ขŒํ‘œ์™€ ๋„ˆ๋น„ ๋ฐ ๋†’์ด๋ฅผ ์ถ”์ •ํ•œ๋‹ค. ๊ทธ๋ฆผ 3์— YOLOv2์˜ ๊ตฌ์กฐ๊ฐ€ ๋‚˜์™€ ์žˆ๋‹ค.

๊ทธ๋ฆผ. 3. YOLOv2 ๋„คํŠธ์›Œํฌ

Fig. 3. YOLOv2 Network

../../Resources/kiee/KIEE.2020.69.2.325/fig3.png

YOLOv3[13]๋Š” YOLOv2๋ฅผ ๊ฐœ์„ ํ•œ ๊ธฐ๋ฒ•์ด๋‹ค. YOLOv2์—์„œ ๊ฒ€์ถœ ๋ถ€๋ถ„์ด ํ•œ ๊ฐ€์ง€ ์Šค์ผ€์ผ์—์„œ๋งŒ ์ˆ˜ํ–‰๋˜๋Š” ๊ฒƒ๊ณผ ๋‹ฌ๋ฆฌ, YOLOv3๋Š” ๊ฒ€์ถœ ๋ถ€๋ถ„์ด ์„ธ ๊ฐ€์ง€ ์Šค์ผ€์ผ์—์„œ ์‹คํ–‰๋˜๊ธฐ ๋•Œ๋ฌธ์—, ๋” ๊นŠ์€ ๋„คํŠธ์›Œํฌ์—์„œ๋„ ํŠน์ง•์„ ํšจ์œจ์ ์œผ๋กœ ๊ฒ€์ถœํ•  ์ˆ˜ ์žˆ๋‹ค. ๋˜ํ•œ, YOLOv2์—์„œ ์ปจ๋ณผ๋ฃจ์…˜ ์ธต์ด 19๊ฐœ์ธ darknet19๋ฅผ ์‚ฌ์šฉํ•œ ๊ฒƒ์— ๋น„ํ•ด YOLOv3์—์„œ๋Š” ๋” ๋†’์€ ์ •ํ™•๋„๋ฅผ ์œ„ํ•ด ์ปจ๋ณผ๋ฃจ์…˜ ์ธต์ด 53๊ฐœ์ธ darknet53์„ ์‚ฌ์šฉํ•œ๋‹ค. ๋„คํŠธ์›Œํฌ ์ธต์ด ๊นŠ์–ด์งˆ์ˆ˜๋ก ๊ธฐ์šธ๊ธฐ ์†์‹ค์˜ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์•„์ง€๊ธฐ ๋•Œ๋ฌธ์— ResNet[15]์—์„œ ์‚ฌ์šฉ๋œ Residual(Shortcut) ๊ธฐ๋ฒ•์„ ์ฐจ์šฉํ•œ๋‹ค. ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ์ž„๋ฒ ๋””๋“œ ์‹œ์Šคํ…œ ์ ์šฉ์„ ์œ„ํ•˜์—ฌ ๊ทœ๋ชจ๊ฐ€ ํฐ YOLOv3 ๋ณด๋‹ค๋Š” YOLOv2๋ฅผ ์ฃผ ๋Œ€์ƒ์œผ๋กœ ๋„คํŠธ์›Œํฌ๋ฅผ ๋ณ€ํ˜•ํ•˜๊ณ  YOLOv3๋Š” ๋น„๊ต๋งŒ์„ ์œ„ํ•ด ์‚ฌ์šฉํ•œ๋‹ค.

4. YOLO ๋„คํŠธ์›Œํฌ ๋ณ€ํ˜•

4.1 ๊ฒ€์ถœ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ์œ„ํ•œ ๊ตฌ์กฐ ๋ณ€๊ฒฝ

์ฒ ํŒ ํ‘œ๋ฉด ๊ฒฐํ•จ์— ๋Œ€ํ•œ ๊ฒ€์ถœ ์„ฑ๋Šฅ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด์„œ YOLOv2์˜ ๊ตฌ์กฐ ๋ณ€๊ฒฝ์„ ์‹œ๋„ํ•œ๋‹ค. Route ๊ธฐ๋Šฅ์€ ์ด์ „ ์ธต์˜ ์ถœ๋ ฅ์„ ์ตœ์ข… ์ถœ๋ ฅ์ธต์— ๊ฒฐํ•ฉ(Concatenation)ํ•˜์—ฌ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ์‹์ด๋‹ค. YOLOv2์˜ ๊ธฐ๋ณธ๊ตฌ์กฐ์—์„œ๋Š” ๊ทธ๋ฆผ 3๊ณผ ๊ฐ™์ด 5๋ฒˆ์งธ ์ปจ๋ณผ๋ฃจ์…˜ ๊ทธ๋ฃน ์ธต๊ณผ ์ตœ์ข… ์ถœ๋ ฅ์ธต์„ ๊ฒฐํ•ฉํ•˜์—ฌ ์‚ฌ์šฉํ•˜๊ณ  ์žˆ๋‹ค. ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ๊ฒฐํ•จ ์ธ์‹๊ณผ์ •์—์„œ์˜ ํŠน์ง• ์†์‹ค์„ ์ค„์ด๊ธฐ ์œ„ํ•ด์„œ ์ด์ „ ์ธต์—์„œ ๋‘ ๊ฐœ์˜ Route๋ฅผ ์—ฐ๊ฒฐํ•˜๋Š” ๋ฐฉ์‹์„ ์‹œ๋„ํ•œ๋‹ค. ๊ทธ๋ฆผ 4์— 4, 5๋ฒˆ์งธ ์ปจ๋ณผ๋ฃจ์…˜ ๊ทธ๋ฃน ์ธต๊ณผ ์ตœ์ข… ์ถœ๋ ฅ์ธต์„ ๊ฒฐํ•ฉํ•˜๋Š” Double Routes ์—ฐ๊ฒฐ ๊ตฌ์กฐ๊ฐ€ ๋‚˜์™€ ์žˆ๋‹ค.

๊ทธ๋ฆผ. 4. Double Routes ์—ฐ๊ฒฐ

Fig. 4. Double Routes connection

../../Resources/kiee/KIEE.2020.69.2.325/fig4.png

4.2 ์ž„๋ฒ ๋””๋“œ ์‹œ์Šคํ…œ ์ˆ˜ํ–‰์„ ์œ„ํ•œ ๋„คํŠธ์›Œํฌ ์ถ•์†Œ

์—ฐ์‚ฐ ์„ฑ๋Šฅ์ด ์ œํ•œ์ ์ธ ์ž„๋ฒ ๋””๋“œ ์‹œ์Šคํ…œ์—์„œ์˜ ๋”ฅ๋Ÿฌ๋‹ ์ˆ˜ํ–‰์„ ์œ„ํ•ด์„œ๋Š” ๋„คํŠธ์›Œํฌ์˜ ์ถ•์†Œ๊ฐ€ ํ•„์š”ํ•˜๋‚˜, ์ด๋Š” ์„ฑ๋Šฅ์˜ ์ €ํ•˜๋ฅผ ๊ฐ€์ ธ์˜จ๋‹ค. ๋”ฐ๋ผ์„œ ์„ฑ๋Šฅ์„ ์ตœ๋Œ€ํ•œ ์œ ์ง€ํ•˜๋ฉด์„œ ์—ฐ์‚ฐ๋Ÿ‰์„ ์ค„์ด๊ธฐ ์œ„ํ•œ ์••์ถ• ๊ธฐ์ˆ ์— ๋Œ€ํ•œ ์—ฐ๊ตฌ๊ฐ€ ์ง„ํ–‰๋˜์–ด ์™”๋‹ค. ๊ฐ€์ค‘์น˜ ๊ฐ€์ง€์น˜๊ธฐ(Weight Pruning), ์–‘์žํ™”/์ด์ง„ํ™”์˜ ์‹œ๋„๋ถ€ํ„ฐ ๋„คํŠธ์›Œํฌ์˜ ๊ตฌ์กฐ๋ฅผ ๋ณ€๊ฒฝํ•˜๋Š” ๋‹ค์–‘ํ•œ ์—ฐ๊ตฌ๊ฐ€ ์ง„ํ–‰๋˜๊ณ  ์žˆ๋‹ค[15]. ๋Œ€์ƒ ๋ฌธ์ œ์— ๋”ฐ๋ผ ๋˜ ๋Œ€์ƒ ๋”ฅ๋Ÿฌ๋‹ ๋„คํŠธ์›Œํฌ์— ๋”ฐ๋ผ ์ ‘๊ทผ๋ฒ•์ด ๋‹ฌ๋ผ์ง€๋ฉฐ, ๊ฒฝ์šฐ์˜ ์ˆ˜๊ฐ€ ๋ฐฉ๋Œ€ํ•œ ๋ฌธ์ œ์ด๋‹ค. ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋น„๊ต์  ๊ฐ„๋‹จํ•œ ์‹œ๋„๋ฅผ ํ†ตํ•ด ๋„คํŠธ์›Œํฌ๊ตฌ์กฐ๋ฅผ ์ถ•์†Œํ•˜๋Š” ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•œ๋‹ค. YOLOv2์—์„œ ๋ฐ˜๋ณต๋˜๋Š” ์ปจ๋ณผ๋ฃจ์…˜ ๊ทธ๋ฃน ์ธต์„ ์ œ๊ฑฐํ•˜์—ฌ ๋„คํŠธ์›Œํฌ ๊ทœ๋ชจ๋ฅผ ์ถ•์†Œํ•œ๋‹ค. ์ฆ‰, ๊ทธ๋ฆผ 5์—์„œ ๋„ค๋ชจ๋กœ ํ‘œ์‹œ๋œ ์ปจ๋ณผ๋ฃจ์…˜ ๊ทธ๋ฃน ์ธต์˜ 4, 5๋ฒˆ์งธ์—์„œ ๋‘ ๋ฒˆ ๋ฐ˜๋ณต์„(2x) ํ•œ๋ฒˆ ๋ฐ˜๋ณต(1x)์œผ๋กœ ๋ณ€๊ฒฝํ•œ๋‹ค. YOLOv3์— ๋Œ€ํ•ด์„œ๋„ ๊ฐ™์€ ๋ฐฉ์‹์„ ์ ์šฉํ•˜์—ฌ ์ถ•์†Œ๋œ ๋ชจ๋ธ์„ ๊ตฌ์„ฑํ•œ๋‹ค.

๊ทธ๋ฆผ. 5. ๋„คํŠธ์›Œํฌ ์ถ•์†Œ

Fig. 5. Network reduction

../../Resources/kiee/KIEE.2020.69.2.325/fig5.png

๊ทธ๋ฆผ. 6. YOLOv2-tiny ์ˆ˜์ •

Fig. 6. Network reduction

../../Resources/kiee/KIEE.2020.69.2.325/fig6.png

4.3 YOLOv2-tiny ๋ชจ๋ธ์˜ ๋ณ€ํ˜•

YOLOv2์™€ v3๋Š” ๊ทœ๋ชจ๊ฐ€ ๋งค์šฐ ํฐ ๋„คํŠธ์›Œํฌ๋กœ ์„ฑ๋Šฅ ์ €ํ•˜๊ฐ€ ์ ์€ ๋ฒ”์œ„๋‚ด์—์„œ ์ผ๋ถ€ ์ถ•์†Œ๋ฅผ ํ•  ๊ฒฝ์šฐ์—๋„ ์ž„๋ฒ ๋””๋“œ ์‹œ์Šคํ…œ์—์„œ ์ˆ˜ํ–‰๋˜๊ธฐ์—๋Š” ์—ฌ์ „ํžˆ ๊ทœ๋ชจ๊ฐ€ ํฌ๋‹ค. ์ถ•์†Œ์œจ์„ ๋†’์ผ ๊ฒฝ์šฐ ์„ฑ๋Šฅ์ด ๊ธ‰๊ฒฉํžˆ ๋–จ์–ด์ง„๋‹ค. ๋”ฐ๋ผ์„œ ์„ฑ๋Šฅ๊ณผ ์ˆ˜ํ–‰ ์†๋„์— ๋Œ€ํ•œ ์‚ฌ์šฉ์ž์˜ ๋‹ค์–‘ํ•œ ์„ ํƒ์„ ์œ„ํ•ด์„œ YOLOv2์˜ ์ถ•์†Œ์™€ ํ•จ๊ป˜ ์‚ฌ์ „์— ๊ฒฝ๋Ÿ‰ํ™”๋œ YOLOv2-tiny[14] ๋ชจ๋ธ๋กœ๋ถ€ํ„ฐ ์—ญ์œผ๋กœ ๊ตฌ์กฐ๋ฅผ ํ™•์žฅ์‹œ์ผœ ์„ฑ๋Šฅ์„ ๋†’์ด๋Š” ์‹œ๋„๋ฅผ ์ˆ˜ํ–‰ํ•œ๋‹ค. YOLOv2-tiny ๋ชจ๋ธ์€ ์ฒ˜๋ฆฌ์†๋„๋Š” ๋งค์šฐ ๋น ๋ฅด๋‚˜ ๊ฒ€์ถœ ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง€๋Š” ๋‹จ์ ์ด ์žˆ๋‹ค. ๊ทธ๋ฆผ 6๊ณผ ๊ฐ™์ด ๋„ค๋ชจ๋กœ ํ‘œ์‹œ๋œ ์ปจ๋ณผ๋ฃจ์…˜ ์ธต(Extra Layers)์„ ์ถ”๊ฐ€ํ•˜์—ฌ YOLOv2-tiny์˜ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•œ๋‹ค.

5. ์‹คํ—˜ ๋ฐ ๊ฒฐ๊ณผ ๋ถ„์„

5.1 ์ž„๋ฒ ๋””๋“œ ๋”ฅ๋Ÿฌ๋‹ ์‹œ์Šคํ…œ ์‹คํ—˜ ํ™˜๊ฒฝ

๊ทธ๋ฆผ. 7. Jetson ๋ณด๋“œ

Fig. 7. Jetson Board

../../Resources/kiee/KIEE.2020.69.2.325/fig7.png

Jetson TX1[12] ๋ณด๋“œ๋Š” ์ž„๋ฒ ๋””๋“œ ํ™˜๊ฒฝ์—์„œ GPU์—ฐ์‚ฐ์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก NVIDIA์‚ฌ์—์„œ ๊ฐœ๋ฐœํ•œ ๊ฐœ๋ฐœ์ž์šฉ ๋„๊ตฌ์ด๋‹ค(๊ทธ๋ฆผ 7). TX1 ๋ณด๋“œ์—๋Š” 256๊ฐœ์˜ ์ฝ”์–ด๋ฅผ ๊ฐ€์ง„ NVIDIA Maxwell GPU๊ฐ€ ๋‚ด์žฅ๋˜์–ด ์žˆ์œผ๋ฉฐ NVIDIA์‚ฌ์—์„œ ์ œ๊ณตํ•˜๋Š” JetPack์„ ์„ค์น˜ํ•˜์—ฌ CUDA, cuDNN ๋ฐ GPU์—ฐ์‚ฐ์— ํ•„์š”ํ•œ ํˆด๋“ค์„ ์†์‰ฝ๊ฒŒ ์„ค์น˜ํ•˜์—ฌ ์‚ฌ์šฉ ํ•  ์ˆ˜ ์žˆ๋‹ค.

5.2 ์‹คํ—˜ ๊ฒฐ๊ณผ

๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ๊ธˆ์†์œผ๋กœ ๋œ ๊ธˆ์† ๋ถ€ํ’ˆ์˜ ํ‘œ๋ฉด์— ๋Œ€ํ•ด North Eastern ๋Œ€ํ•™์—์„œ ๊ตฌ์„ฑํ•œ NEU surface ๋ฐ์ดํ„ฐ[11]๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‹คํ—˜์„ ์ง„ํ–‰ํ•œ๋‹ค.

ํ‘œ 1. ์‹คํ—˜์— ์‚ฌ์šฉ๋œ ๋ฐ์ดํ„ฐ ์ˆ˜

Table 1. Number of data used for experiments

No. of training data

No. of test data

NEU surface

1,440

360

ํ‘œ 2. ํ•™์Šต ํŒŒ๋ผ๋ฏธํ„ฐ

Table 2. Parameters for training

Learning rate

Burn in

Batch size

Max batches

0.001

1000

64

24000

์‹คํ—˜์— ์‚ฌ์šฉ๋œ ๋ฐ์ดํ„ฐ๋Š” ๊ฐ ํด๋ž˜์Šค๋‹น 300์žฅ ์”ฉ ์ด 6๊ฐœ์˜ ํด๋ž˜์Šค๋กœ ๊ตฌ์„ฑ์ด ๋˜์–ด์žˆ์œผ๋ฉฐ, ํ•™์Šต๊ณผ ํ…Œ์ŠคํŠธ์˜ ๊ตฌ์„ฑ์€ ํ‘œ 1๊ณผ ๊ฐ™์œผ๋ฉฐ, ํ•™์Šต์— ์‚ฌ์šฉ๋œ ํŒŒ๋ผ๋ฏธํ„ฐ๋Š” ํ‘œ 2์™€ ๊ฐ™๋‹ค.

YOLOv2์™€ v3์—์„œ ์•ต์ปค๋ฐ•์Šค ์ตœ์ ํ™”[7]๋Š” ๊ฒฐํ•จ๋“ค์˜ ๋ผ๋ฒจ๋“ค์„ ๋น„๊ตํ•˜์—ฌ k๊ฐœ์˜ ๋ฐ•์Šค๋“ค๋กœ ๊ตฐ์ง‘ํ™” ํ•˜์—ฌ ํ•ด๋‹น ๋ฐ์ดํ„ฐ์˜ ๊ฒฐํ•จํ•ฉ ๋ถ„ํฌ์— ๋งž๊ฒŒ ์•ต์ปค๋“ค์„ ๊ตฌ์„ฑํ•œ๋‹ค. ๊ทธ๋ฆผ 8์—์„œ ์ขŒ์ธก์˜ Anchor 5๋Š” YOLOv2 ์—์„œ์˜ ์ตœ์ ํ™” ๊ฒฐ๊ณผ์ด๊ณ , Anchor 9 ์€ YOLOv3์—์„œ์˜ ์ตœ์ ํ™” ๊ฒฐ๊ณผ๋ฅผ ๋‚˜ํƒ€๋‚ธ๋‹ค.

๊ทธ๋ฆผ. 8. ์•ต์ปค๋ฐ•์Šค ์ตœ์ ํ™”, Anchor 5 (v2), Anchor 9 (v3)

Fig. 8. Optimization of anchor box, Anchor 5 (v2), Anchor 9 (v3)

../../Resources/kiee/KIEE.2020.69.2.325/fig8.png

YOLOv2์™€ v3 ๋ฐ ์ œ์•ˆ๋œ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ณ€ํ˜•๋œ ๋ชจ๋ธ์— ๋Œ€ํ•ด์„œ ๊ฒ€์ถœ ์„ฑ๋Šฅ๊ณผ BFLOPS(Billion Flops) ๋ฐ fps๋ฅผ ๋น„๊ตํ•œ ๊ฒฐ๊ณผ๊ฐ€ ํ‘œ 3์— ๋‚˜์™€ ์žˆ๋‹ค. YOLOv2-route๋Š” 4.2์ ˆ์—์„œ ์„ค๋ช…ํ•œ Double Routes ์—ฐ๊ฒฐ ๊ตฌ์กฐ์ด๊ณ , YOLOv2_r 4.2์ ˆ์—์„œ ์–ธ๊ธ‰ํ•œ ์ถ•์†Œ ๋ชจ๋ธ์ด๋‹ค. YOLOv2_route_r์€ YOLOv2-route๋ฅผ ์ถ•์†Œํ•œ ๊ฒƒ์ด๊ณ , YOLOv3_r์€ YOLOv3์˜ ์ถ•์†Œ ๋ชจ๋ธ์ด๋‹ค. YOLOv2_tiny_m์€ YOLOv2_tiny๋ฅผ ์ˆ˜์ •ํ•œ ๋ชจ๋ธ์ด๋‹ค(4.3์ ˆ ์ฐธ๊ณ ).

ํ‘œ 3. ์ˆ˜์น˜ ์„ฑ๋Šฅ ๊ฒฐ๊ณผ

Table 3. Numerical performance results

mAP(%)

BFLOPS

fps

YOLOv2

65.14

29.362

4.8

YOLOv2-route

67.04

32.640

4.2

YOLOv3

58.20

65.326

2.5

YOLOv2_r

65.53

25.818

5.1

YOLOv2_route_r

64.41

29.096

4.6

YOLOv3_r

55.78

46.985

3.3

YOLOv2-tiny

36.77

6.953

12.5

YOLOv2_tiny_m

62.33

13.280

8.8

YOLOv2-route๋Š” ์›๋ณธ์˜ ๊ฒ€์ถœ ์„ฑ๋Šฅ์ธ 65.14% ๋ณด๋‹ค 2% ์ฆ๊ฐ€๋œ 67.04%์˜ ํ–ฅ์ƒ์„ ๋ณด์˜€์œผ๋‚˜ ์—ฐ์‚ฐ๋Ÿ‰์ธ BFLOPS๋Š” 32.64๋กœ ์ฆ๊ฐ€ํ•˜๊ณ  fps๋Š” 4.2๋กœ ๊ฐ์†Œํ•˜์˜€๋‹ค. ์„ฑ๋Šฅ ๊ฐœ์„ ์„ ์œ„ํ•ด ๋” ๊ทœ๋ชจ๊ฐ€ ์ปค์ง„ YOLOv3๋Š” ์ฒ ํŒ ๊ฒฐํ•จ ๊ฒ€์ถœ ๋ฌธ์ œ์—๋Š” ์˜คํžˆ๋ ค 58.2%๋กœ ๊ฒ€์ถœ ์„ฑ๋Šฅ์ด ๋–จ์–ด์กŒ๋‹ค. ๊ทธ๋ฆฌ๊ณ  fps๋Š” YOLOv2์˜ ์ ˆ๋ฐ˜ ์ˆ˜์ค€์ธ 2.5๋ฅผ ๊ธฐ๋กํ–ˆ๋‹ค. YOLOv2_r ์ถ•์†Œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์€ 65.53%๋กœ ์›๋ณธ๋ณด๋‹ค 0.5% ํ–ฅ์ƒ๋˜์—ˆ๊ณ  fps๋„ 5.1 ๋กœ ์†Œํญ ์ฆ๊ฐ€๋˜์—ˆ๋‹ค. YOLOv2_route ๋ชจ๋ธ์„ ์ถ•์†Œํ•œ YOLOv2_route_r์—์„œ๋Š” ์„ฑ๋Šฅ์€ 64.41%๋กœ ์•ฝ๊ฐ„ ์ €ํ•˜๋˜์—ˆ๊ณ  ์ฒ˜๋ฆฌ์†๋„๋Š” 4.6 fps๋กœ ์ถ•์†Œ ์ „๋ณด๋‹ค๋Š” ์ฆ๊ฐ€๋˜์—ˆ์œผ๋‚˜ ๊ธฐ๋ณธ YOLOv2์— ๋น„ํ•ด์„œ๋Š” ๋‘ ๊ฐ€์ง€ ์ธก๋ฉด ๋ชจ๋‘ ์•ฝ๊ฐ„์”ฉ ๋ชป ๋ฏธ์ณค๋‹ค. YOLOv3_r๋„ ์ฒ˜๋ฆฌ์†๋„๋Š” ์›๋ณธ์— ๋น„ํ•ด ์•ฝ๊ฐ„ ๋†’์•„์กŒ์œผ๋‚˜ ์„ฑ๋Šฅ์ด ์ €ํ•˜๋˜์–ด ๊ฒฐํ•จ ๋ฌธ์ œ์— ๋Œ€ํ•ด์„œ๋Š” ํฐ ํ๋ชจ์˜ ๋„คํŠธ์›Œํฌ๊ฐ€ ์œ ์šฉํ•˜์ง€ ๋ชปํ•จ์„ ๋ณด์—ฌ์ค€๋‹ค. ์ •๋ฆฌํ•˜๋ฉด YOLOv2 ๋ณด๋‹ค ๋‘ ์ง€ํ‘œ์—์„œ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๊ฒƒ์€ YOLOv2_r ์ž„์„ ์•Œ ์ˆ˜ ์žˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ 5.1 fps๋Š” ์—ฌ์ „ํžˆ ๋‚ฎ์€ ์ฒ˜๋ฆฌ ์†๋„์ด๊ธฐ ๋•Œ๋ฌธ์—, ์ด๋ฅผ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด์„œ ์†๋„๊ฐ€ ๋น ๋ฅธ YOLOv2-tiny์˜ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜์—ฌ YOLOv2_tiny_m์„ ์ œ์•ˆํ•˜์˜€๊ณ , 62.33% ์„ฑ๋Šฅ์— 8.8 fps๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์—ˆ๋‹ค. ์ด๊ฒƒ์€ YOLOv2์— ๋น„ํ•ด์„œ ์„ฑ๋Šฅ์€ 2.81%๋งŒ ๊ฐ์†Œํ•˜๋ฉด์„œ ์ฒ˜๋ฆฌ์†๋„๋Š” ๊ฑฐ์˜ 2๋ฐฐ ๊ฐ€๊นŒ์ด ํ–ฅ์ƒ ์‹œํ‚จ ๊ฐœ์„ ๋œ ๊ฒฐ๊ณผ์ด๋‹ค.

ํ‘œ 1์— ๋‚˜์˜จ YOLO ๋ณ€ํ˜• ๋ชจ๋ธ์— ๋Œ€ํ•œ ์„ฑ๋Šฅ ๋น„๊ต ๊ทธ๋ž˜ํ”„๊ฐ€ ๊ทธ๋ฆผ 8์— ๋‚˜์™€ ์žˆ๋‹ค. ์„ฑ๋Šฅ mAP์™€ ์—ฐ์‚ฐ๋Ÿ‰ BFLOPS์— ๋Œ€ํ•œ ์ˆ˜์น˜๋ฅผ ๋„์‹ํ™” ํ•œ ๊ฒƒ์œผ๋กœ ์ขŒ์ƒ์— ์œ„์น˜ํ• ์ˆ˜๋ก ์šฐ์ˆ˜ํ•จ์„ ๋‚˜ํƒ€๋‚ธ๋‹ค. YOLOv2_tiny_m ๋ชจ๋ธ์ด ๋‘ ๊ฐ€์ง€ ์ง€ํ‘œ๋ฅผ ๋งŒ์กฑ์‹œํ‚ด์„ ์•Œ ์ˆ˜ ์žˆ๋‹ค. ๊ทธ๋ฆผ 9์—๋Š” mAP์™€ fps์— ๋Œ€ํ•œ ๊ทธ๋ž˜ํ”„๊ฐ€ ๋‚˜์™€ ์žˆ์œผ๋ฉฐ ์šฐ์ƒ์— ์œ„์น˜ํ• ์ˆ˜๋ก ์šฐ์ˆ˜ํ•จ์„ ๋‚˜ํƒ€๋‚ธ๋‹ค. ์—ญ์‹œ YOLOv2_tiny_m ๋ชจ๋ธ์ด ๊ฐ€์žฅ ์ข‹์€ ๋Œ€์•ˆ์ž„์„ ๋ณด์—ฌ์ฃผ๊ณ  ์žˆ๋‹ค.

ํ˜„์žฌ๋Š” ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง€๋Š” Jetson TX1 ๋ณด๋“œ์—์„œ ์‹คํ—˜ํ•˜์˜€์œผ๋‚˜ ์—ฐ์‚ฐ ์†๋„๊ฐ€ ๋น ๋ฅธ Jetson TX2 ๋ณด๋“œ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด, ๊ฐ™์€ ๊ฒ€์ถœ ์„ฑ๋Šฅ์— ๋” ๋†’์€ fps๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์–ด์„œ, ์‹ค์ œ ํ˜„์žฅ ์ ์šฉ์˜ ๊ฐ€๋Šฅ์„ฑ์„ ๊ธฐ๋Œ€ํ•  ์ˆ˜ ์žˆ๋‹ค.

๊ทธ๋ฆผ. 9. YOLO ๋ณ€ํ˜• ๋ชจ๋ธ ์„ฑ๋Šฅ ๋น„๊ต 1 (mAP ๋Œ€ BFLOPS)

Fig. 9. Performance comparison 1 of various YOLO Networks (mAP vs BFLOPS)

../../Resources/kiee/KIEE.2020.69.2.325/fig9.png

๊ทธ๋ฆผ. 10. YOLO ๋ณ€ํ˜• ๋ชจ๋ธ ์„ฑ๋Šฅ ๋น„๊ต 2 (mAP ๋Œ€ fps)

Fig. 10. Performance comparison 2 of various YOLO Networks (mAP vs fps)

../../Resources/kiee/KIEE.2020.69.2.325/fig10.png

6. ๊ฒฐ ๋ก 

๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ์ฒ ํŒ ํ‘œ๋ฉด ๊ฒฐํ•จ์— ๋Œ€ํ•œ ๋”ฅ๋Ÿฌ๋‹ ๊ธฐ๋ฐ˜ ๊ฒ€์ถœ ๊ธฐ๋ฒ•์„ ์ ์šฉํ•˜์—ฌ ๊ฒ€์ถœ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ณ , ์ž„๋ฒ ๋””๋“œ ์‹œ์Šคํ…œ ์ ์šฉ์„ ๊ฒ€ํ† ํ•˜๊ธฐ ์œ„ํ•ด์„œ YOLO ๋„คํŠธ์›Œํฌ์˜ ๋ณ€ํ˜•์„ ์‹œ๋„ํ•˜์˜€๋‹ค. NEU ์ฒ ํŒ ๋ฐ์ดํ„ฐ๋ฅผ ๋Œ€์ƒ์œผ๋กœ Jetson TX1 ๋ณด๋“œ์˜ ์ž„๋ฒ ๋””๋“œ ํ™˜๊ฒฝ์—์„œ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•˜์˜€์œผ๋ฉฐ, 8๊ฐ€์ง€์˜ ๋ชจ๋ธ์— ๋Œ€ํ•ด์„œ ๊ฒ€์ถœ ์„ฑ๋Šฅ๊ณผ ์ฒ˜๋ฆฌ์†๋„๋ฅผ ๋น„๊ตํ•˜์˜€๋‹ค. ์–ป์–ด์ง„ YOLOv2_tiny_m ๋ชจ๋ธ์€ ๊ฒ€์ถœ ์„ฑ๋Šฅ๊ณผ fps ์ˆ˜์น˜์—์„œ ๋‹ค๋ฅธ ๋ชจ๋ธ์— ๋น„ํ•ด์„œ ์˜๋ฏธ ์žˆ๋Š” ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ฃผ์—ˆ๋‹ค. ํ–ฅํ›„ ์ฒ˜๋ฆฌ ์†๋„ ํ–ฅ์ƒ์„ ์œ„ํ•œ ๋‹ค์–‘ํ•œ ๋„คํŠธ์›Œํฌ ์ถ•์†Œ ๋˜๋Š” ์••์ถ• ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ๋„์ž…์ด ํ•„์š”ํ•˜๊ณ , ๋‹ค๋ชฉ์  ์ตœ์ ํ™” ๋ถ„์•ผ์—์„œ ๋‹ค๋ฃจ์–ด์ง€๋Š” ์„ฑ๋Šฅ๊ณผ ์ฒ˜๋ฆฌ์†๋„์— ๋Œ€ํ•œ ๋‹ค์–‘ํ•œ ํŒŒ๋ ˆํ† (Pareto) ํ•ด๋ฅผ ์ œ๊ณตํ•  ํ•„์š”๊ฐ€ ์žˆ๋‹ค.

References

1 
Y. Park, I. S. Kweon, 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 
S. Ghorai, A. Mukherjee, M. Gangadaran, P. K. Dutta, 2013, Automatic Defect Detection on Hot-Rolled Flat Steel Products, IEEE Trans. on Instrumentation and Measurement, Vol. 62, No. 3, pp. 612-621DOI
3 
J.-K. Park, N. Kwon, J.-H. Park, D. Kang, 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
4 
B.-K. Kwon, J.-S. Won, D.-J. Kang, 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
5 
H. Choi, K. Seo, 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 
R. Girshick, J. Donahue, T. Darrell, J. Malik, 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
7 
Shaoqing Ren, Kaiming He, Ross Girshick, Jian Sun, 2015, Faster R-CNN: Towards Real-time Object Detection with Region Proposal Networks, Advances in Neural Information Processing Systems, pp. 91-99DOI
8 
H. Choi, K. Seo, 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
9 
Joseph Redmon, Santosh Divvala, Ross Girshick, Ali Farhadi, 2016, You Only Look Once: Unified, Real-Time Object Detection, in Proceedings of The IEEE Conference on Computer Vision and Pattern RecognitionDOI
10 
M. Lee, K. Seo, 2018, Comparison of CNN Methods for Defects Detection, in Proceedings of Information and Control Symposium ICS'2018, pp. 101-102Google Search
11 
K. Song, Y. Yan, NEU Surface Defect Database, http://faculty.neu.edu.cn/yunhyan/NEU_surface_defect_database.htmlGoogle Search
12 
Jetson TX1 Module, https://developer.nvidia.com/embedded/jetson-tx1Google Search
13 
Joseph Redmon, Ali Farhadi, 2018, YOLOv3: An Incremental Improvement, arXiv preprint arXiv:1804.02767Google Search
14 
Tiny YOLOv2, https://gallery.azure.ai/Model/Tiny-YOLOv2Google Search
15 
Y. J. Lee, Y. H. Moon, J. Y. Park, O. G. Min, 2019, Recent R&D Trends for Lightweight Deep Learning, ETRI Electronics and Telecommunications Trends, Vol. 34, No. 6, pp. 40-50Google Search

์ €์ž์†Œ๊ฐœ

์ด๊ฑด์˜ (Keon Young Yi )
../../Resources/kiee/KIEE.2020.69.2.325/au1.png

1982๋…„ ํ•œ์–‘๋Œ€ ์ „๊ธฐ๊ณตํ•™๊ณผ ์กธ์—…. 1984๋…„ ๋™ ๋Œ€ํ•™์› ์ „๊ธฐ๊ณตํ•™๊ณผ ์กธ์—…(์„์‚ฌ).

1993๋…„ ๋™ ๋Œ€ํ•™์› ์ „๊ธฐ๊ณตํ•™๊ณผ ์กธ์—…(๊ณต๋ฐ•).

1994~1996๋…„ The Ohio State University, Dept. of EE, ์—ฐ๊ตฌ์›

2004~ 2005๋…„ University of Hawaii, Dept. of ME, ๋ฐฉ๋ฌธ์—ฐ๊ตฌ์›

2014๋…„~ํ˜„์žฌ ๋ฏธ๋‹ˆ๋“œ๋ก ์ž์œจ๋น„ํ–‰ ๊ฒฝ์ง„๋Œ€ํšŒ ์กฐ์ง์œ„์›์žฅ

1996๋…„~ํ˜„์žฌ ๊ด‘์šด๋Œ€ ์ „๊ธฐ๊ณตํ•™๊ต ๊ต์ˆ˜. ๊ด€์‹ฌ๋ถ„์•ผ๋Š” ๋“œ๋ก  ์ œ์–ด, ๋“œ๋ก  ์ธ์‹

์ •์„ ์žฌ (Sunjae Jeong)
../../Resources/kiee/KIEE.2020.69.2.325/au2.png

2020๋…„ ์„œ๊ฒฝ๋Œ€ํ•™๊ต ์ „์ž๊ณตํ•™๊ณผ ์กธ์—…(ํ•™์‚ฌ).

2022๋…„ ์„œ๊ฒฝ๋Œ€ํ•™๊ต ๋Œ€ํ•™์› ์ „์ž์ปดํ“จํ„ฐ๊ณตํ•™๊ณผ ์กธ์—…์˜ˆ์ •(์„์‚ฌ).

๊ด€์‹ฌ๋ถ„์•ผ๋Š” ์ง„ํ™”์—ฐ์‚ฐ, ๋จธ์‹ ๋น„์ „, ๋”ฅ๋Ÿฌ๋‹.

์„œ๊ธฐ์„ฑ (Kisung Seo)
../../Resources/kiee/KIEE.2020.69.2.325/au3.png

1986๋…„ ์—ฐ์„ธ๋Œ€ํ•™๊ต ์ „๊ธฐ๊ณตํ•™๊ณผ ์กธ์—…(๊ณตํ•™์‚ฌ)

1988๋…„ ์—ฐ์„ธ๋Œ€ํ•™๊ต ๋Œ€ํ•™์› ์ „๊ธฐ๊ณตํ•™๊ณผ ์กธ์—…(์„์‚ฌ).

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„์•ผ๋Š” ์ง„ํ™”์—ฐ์‚ฐ, ๋”ฅ๋Ÿฌ๋‹, ๋จธ์‹ ๋น„์ „, ๊ธฐ์ƒ์˜ˆ์ธก, ์ง€๋Šฅ๋กœ๋ด‡