License Plate Detection Application

Overview

LicensePlate_Project ๐Ÿš— ๐Ÿš™

[Project] 2021.02 ~ 2021.09 License Plate Detection Application

Overview


1. ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘ ๋ฐ ๋ผ๋ฒจ๋ง

์ฐจ๋Ÿ‰ ๋ฒˆํ˜ธํŒ ์ด๋ฏธ์ง€๋ฅผ ์ง์ ‘ ์ˆ˜์ง‘ํ•˜์—ฌ ๊ฐ ์ด๋ฏธ์ง€์— ๋Œ€ํ•ด '๋ฒˆํ˜ธํŒ ๊ธ€์ž'์™€ '๋ฒˆํ˜ธํŒ ๋„ค ๊ผญ์ง“์ ์˜ x,y ์ขŒํ‘œ'๋ฅผ ๋ผ๋ฒจ๋ง ํ•œ๋‹ค.

๋ฒˆํ˜ธํŒ ์ด๋ฏธ์ง€
๋ผ๋ฒจ๋ง 20210210_222919.jpg 1481 2773 2043 2689 2043 2794 1486 2883 36์กฐ 2428

ํ…์ŠคํŠธ ํŒŒ์ผ๋กœ ์ €์žฅ๋œ ๋ผ๋ฒจ๋ง ์ •๋ณด๋Š” ๋ฒˆํ˜ธํŒ ๋„ค ๊ผญ์ง“์ ์˜ ์ ˆ๋Œ€ ์ขŒํ‘œ์™€ ๋ฒˆํ˜ธํŒ ๊ธ€์ž๋ฅผ ํฌํ•จํ•˜๊ณ  ์žˆ๋‹ค. ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ 20%๋ฅผ ๊ฒ€์ฆ ๋ฐ์ดํ„ฐ๋กœ ๋‚˜๋ˆ„์–ด ๋ฐ์ดํ„ฐ์…‹ ์ค€๋น„๋ฅผ ๋งˆ์นœ๋‹ค. ์ตœ์ข… ๋ฐ์ดํ„ฐ์…‹ ๊ตฌ์„ฑ์€ ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค.

ํ•™์Šต ๋ฐ์ดํ„ฐ ๊ฒ€์ฆ ๋ฐ์ดํ„ฐ
1635์žฅ 409์žฅ

2. YOLOv5 ํ•™์Šต (Pytorch-YOLOv5)

  • ์ฐธ๊ณ : https://github.com/ultralytics/yolov5

  • ์ธํ’‹ ๋ฐ์ดํ„ฐ ์ค€๋น„
    ์›๋ณธ ์ด๋ฏธ์ง€๋Š” ๋ฒˆํ˜ธํŒ ์˜์—ญ์„ ํƒ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ๊ณง์žฅ YOLO์˜ ์ž…๋ ฅ์œผ๋กœ ์‚ฌ์šฉ๋˜๊ธฐ ๋•Œ๋ฌธ์—, YOLO์˜ ์ž…๋ ฅ ํ˜•์‹์— ๋งž์ถ”๊ธฐ ์œ„ํ•ด ๊ฐ ์ด๋ฏธ์ง€ ๋งˆ๋‹ค ์ด๋ฏธ์ง€ ํŒŒ์ผ๋ช…๊ณผ ๋™์ผํ•œ ์ด๋ฆ„์˜ ํ…์ŠคํŠธ ํŒŒ์ผ์„ ๋งŒ๋“ค์–ด bounding box์˜ ์ขŒํ‘œ ์ •๋ณด๋ฅผ class, x_center, y_center, width, height์˜ ํฌ๋งท์˜ ๋ฌธ์ž์—ด๋กœ ์ €์žฅํ•œ๋‹ค. ์ด ๋•Œ, class๋ฅผ ์ œ์™ธํ•œ ๋‚˜๋จธ์ง€ ๊ฐ’์€ ๋ชจ๋‘ 0-1 ์‚ฌ์ด์˜ ์ƒ๋Œ€ ์ขŒํ‘œ๋กœ ๋ณ€ํ™˜ํ•œ๋‹ค.

โ”œโ”€โ”€ Yolo_input
    โ”œโ”€โ”€ train
    โ”‚   โ”œโ”€โ”€ images
    โ”‚   โ”‚   โ”œโ”€โ”€ 1.jpg
    โ”‚ 	โ”‚   โ”œโ”€โ”€ 2.jpg
    โ”‚ 	โ”‚  	โ”‚     :
    โ”‚ 	โ”‚  		  
    โ”‚   โ”œโ”€โ”€ labels
    โ”‚	    โ”œโ”€โ”€ 1.txt
    โ”‚	    โ”œโ”€โ”€ 2.txt
    โ”‚	   	โ”‚     :
    โ”‚	
    โ””โ”€โ”€ val
 	    โ”œโ”€โ”€ images
 	    โ”œโ”€โ”€ labels
  • dataset.yaml ์ค€๋น„
    Custom ๋ฐ์ดํ„ฐ์…‹์— YOLOv5 ํ•™์Šต ์ฝ”๋“œ๋ฅผ ๊ทธ๋Œ€๋กœ ์“ธ ๊ฒƒ์ด๊ธฐ ๋•Œ๋ฌธ์—, ๋ฐ์ดํ„ฐ์…‹ ์„ธํŒ… ๋ถ€๋ถ„๋งŒ ์ˆ˜์ •ํ•œ๋‹ค. dataset.yaml ํŒŒ์ผ์— ํ•™์Šต, ๊ฒ€์ฆ ๋ฐ์ดํ„ฐ ๊ฒฝ๋กœ์™€ ๊ฐ์ฒด ํด๋ž˜์Šค ์ •๋ณด๋ฅผ ๊ธฐ์ž…ํ•œ๋‹ค. ์šฐ๋ฆฌ ํ”„๋กœ์ ํŠธ์˜ ๊ฒฝ์šฐ ํƒ์ง€ํ•˜๋Š” ๊ฐ์ฒด๊ฐ€ ์ฐจ๋Ÿ‰ ๋ฒˆํ˜ธํŒ ํ•˜๋‚˜์ด๋ฏ€๋กœ ํด๋ž˜์Šค ๋ผ๋ฒจ์„ 0์œผ๋กœ, ์ด๋ฆ„์„ 'plate' ๋กœ ํ•œ๋‹ค.

  • YOLO ๋ชจ๋ธ ์„ ํƒ
    ๋ณธ ํ”„๋กœ์ ํŠธ๋ฅผ ์œ„ํ•ด ๊ฐ€์žฅ ์ž‘๊ณ  ๋น ๋ฅธ ๋ชจ๋ธ์ธ YOLOv5s๋ฅผ ์‚ฌ์šฉํ•˜์˜€๋‹ค.


3. ๊ผญ์ง“์  ์˜ˆ์ธก ๋ชจ๋ธ ํ•™์Šต

  • ์‚ฌ์šฉํ•œ ๋ชจ๋ธ : timm์œผ๋กœ ์‚ฌ์ „ํ•™์Šต๋œ Resnet18 ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์˜€๋‹ค

  • ์ฒซ ๋ฒˆ์งธ ๋ฐฉ๋ฒ•

    1. ์‚ฌ์šฉ๋œ ์ด๋ฏธ์ง€ : ๋„ค ๊ผญ์ง“์  ์ขŒํ‘œ๊ฐ’์„ ์ด์šฉํ•˜์—ฌ ๋งŒ๋“  ๋ฐ”์šด๋”ฉ ๋ฐ•์Šค์—์„œ ๊ฐ ์ถ•์œผ๋กœ 1%์”ฉ ๋Š˜์ธ ์ด๋ฏธ์ง€

    2. ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•: ์ „๋‹จ ๋ณ€ํ™˜(shear transformation), ์‚ฌ์ง„ํ•ฉ์„ฑ, ๋ฐ๊ธฐ์กฐ์ ˆ, ๋ฆฌ์‚ฌ์ด์ฆˆ
      ์ž…๋ ฅ ์ด๋ฏธ์ง€๋ฅผ ์ „๋‹จ ๋ณ€ํ™˜ ๊ธฐ๋ฒ•์„ ์ด์šฉํ•ด x, y์ถ•์œผ๋กœ ๋žœ๋คํ•˜๊ฒŒ ๋ณ€ํ™˜ํ•˜๋ฉด ๊ฒ€์€์ƒ‰ ์—ฌ๋ฐฑ ๋ถ€๋ถ„์ด ์ƒ๊ฒจ, ์ด ๋ถ€๋ถ„์„ ๋‹ค๋ฅธ ์ด๋ฏธ์ง€์—์„œ ๋žœ๋คํ•˜๊ฒŒ ๊ฐ€์ ธ์™€ ํ•ฉ์„ฑ์‹œ์ผฐ๋‹ค. ์ด ์ด๋ฏธ์ง€์— ๋žœ๋ค์œผ๋กœ ๋ฐ๊ธฐ์กฐ์ ˆ์„ ์ถ”๊ฐ€ํ•˜์—ฌ, 128x128 ์ด๋ฏธ์ง€๋กœ ๋ฆฌ์‚ฌ์ด์ฆˆํ•œ ์ด๋ฏธ์ง€๋ฅผ ๋ชจ๋ธ์— ์ž…๋ ฅ์œผ๋กœ ๋„ฃ์—ˆ๋‹ค.

    3. ๋ฌธ์ œ์  : ๊ฒ€์€์ƒ‰ ๋ถ€๋ถ„์„ ๋‹ค๋ฅธ ์‚ฌ์ง„์œผ๋กœ ํ•ฉ์„ฑ์‹œ์ผฐ๋”๋‹ˆ ์‹ค์„ธ๊ณ„ ๋ฐ์ดํ„ฐ์™€ ๊ดด๋ฆฌ๊ฐ์ด ์ƒ๊ฒจ ์„ฑ๋Šฅ ์ €ํ•˜ ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•˜์˜€๋‹ค.

    ์‚ฌ์šฉ๋œ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ์ฆ๊ฐ•1 ๋ฐ์ดํ„ฐ์ฆ๊ฐ•2
    ์‚ฌ์šฉ๋œ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ์ฆ๊ฐ•1 ๋ฐ์ดํ„ฐ์ฆ๊ฐ•2
  • ๋‘ ๋ฒˆ์งธ ๋ฐฉ๋ฒ•

    1. ์‚ฌ์šฉ๋œ ์ด๋ฏธ์ง€ : ์›๋ณธ ์ด๋ฏธ์ง€

    2. ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•: ์ „๋‹จ ๋ณ€ํ™˜, ๋ฐ๊ธฐ์กฐ์ ˆ, ๋ฆฌ์‚ฌ์ด์ฆˆ ์ž…๋ ฅ ์ด๋ฏธ์ง€์™€ ๋ผ๋ฒจ๋ง์„ ํ†ตํ•ด ์•Œ๋ ค์ง„ ๋ฒˆํ˜ธํŒ ๊ผญ์ง“์ ์˜ ์ขŒํ‘œ๋“ค์„ ์ „๋‹จ ๋ณ€ํ™˜ ๊ธฐ๋ฒ•์„ ์ด์šฉํ•ด ๋žœ๋ค ๊ฐ’์œผ๋กœ ๋ณ€ํ™˜ํ•œ๋‹ค. ์ด ์ด๋ฏธ์ง€์—์„œ ๋ฒˆํ˜ธํŒ์˜ ์ขŒํ‘œ๋ฅผ ๊ธฐ์ค€์œผ๋กœ margin์„ ์ฃผ๊ณ , ๊ทธ ์ง€์ ์œผ๋กœ๋ถ€ํ„ฐ ๋žœ๋คํ•˜๊ฒŒ ์ขŒํ‘œ๋ฅผ ์ฐ์–ด ์ด๋ฏธ์ง€๋ฅผ ์ž๋ฅธ ๊ฒƒ์„ ์‚ฌ์šฉ. ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ์ฒซ ๋ฒˆ์งธ ๋ฐฉ๋ฒ•์—์„œ ๋‚˜์™”๋˜ ๊ฒ€์€ ์—ฌ๋ฐฑ ๋ถ€๋ถ„์ด ๋‚˜์˜ค์ง€ ์•Š์œผ๋ฏ€๋กœ ์‹ค์„ธ๊ณ„ ๋ฐ์ดํ„ฐ์™€ ๋” ๊ทผ์ ‘ํ•˜๋‹ค. ์ด ์ด๋ฏธ์ง€์— ๋žœ๋ค์œผ๋กœ ๋ฐ๊ธฐ์กฐ์ ˆ์„ ์ถ”๊ฐ€ํ•˜์—ฌ, 128x128 ์ด๋ฏธ์ง€๋กœ ๋ฆฌ์‚ฌ์ด์ฆˆํ•œ ์ด๋ฏธ์ง€๋ฅผ ๋ชจ๋ธ์— ์ž…๋ ฅ์œผ๋กœ ๋„ฃ์—ˆ๋‹ค.

    ์‚ฌ์šฉ๋œ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ์ฆ๊ฐ•1 ๋ฐ์ดํ„ฐ์ฆ๊ฐ•2
    ์‚ฌ์šฉ๋œ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ์ฆ๊ฐ•1 ๋ฐ์ดํ„ฐ์ฆ๊ฐ•2
  • Output : ์ƒํ•˜์ขŒ์šฐ ๋„ค ๊ผญ์ง“์ ์— ๋Œ€ํ•œ X,Y ์ƒ๋Œ€์ขŒํ‘œ


4. ๊ธ€์ž ์˜ˆ์ธก ๋ชจ๋ธ ํ•™์Šต

  • ์‚ฌ์šฉํ•œ ๋ชจ๋ธ : timm์œผ๋กœ ์‚ฌ์ „ํ•™์Šต๋œ Resnet18 ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์˜€๋‹ค.

  • ์ฒซ ๋ฒˆ์งธ ๋ฐฉ๋ฒ•

    1. ์‚ฌ์šฉ๋œ ์ด๋ฏธ์ง€ : ์›๋ณธ ์ด๋ฏธ์ง€์˜ ๋„ค ๊ผญ์ง“์  ์ขŒํ‘œ์— ๋Œ€ํ•œ ground truth๋ฅผ ์ด์šฉํ•˜์—ฌ (128, 256)์˜ ํฌ๊ธฐ๋กœ ํˆฌ์˜๋ณ€ํ™˜ํ•œ ์ด๋ฏธ์ง€

    2. ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•: Salt & Pepper ๋…ธ์ด์ฆˆ ์‹ค์ œ ์ฐจ๋Ÿ‰์˜ ๋ฒˆํ˜ธํŒ์€ ๋จผ์ง€ ๋ฐ ๋ฒŒ๋ ˆ์™€ ๊ฐ™์€ ์ด๋ฌผ์งˆ ๋•Œ๋ฌธ์— ์–ผ๋ฃฉ๋œ๋ฃฉํ•œ ๊ฒฝ์šฐ๊ฐ€ ๋งŽ๋‹ค. ๋”ฐ๋ผ์„œ ์ž…๋ ฅ ๋ฐ์ดํ„ฐ์— ๋žœ๋คํ•œ ๋…ธ์ด์ฆˆ๋ฅผ ์ถ”๊ฐ€ํ•˜์—ฌ ์ผ๋ฐ˜์ ์ธ ์ƒํ™ฉ๊นŒ์ง€ ์ปค๋ฒ„ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜์˜€๋‹ค.

    3. ๋ฌธ์ œ์  : ์‹ค์ œ ์ถ”๋ก  ๊ณผ์ •์—์„œ๋Š” ๊ผญ์ง“์  ์˜ˆ์ธก ๋ชจ๋ธ๋กœ๋ถ€ํ„ฐ ์˜ˆ์ธก๋œ ๊ผญ์ง“์  ๊ฐ’์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ •๋ ฌ๋œ ๋ฒˆํ˜ธํŒ ์ด๋ฏธ์ง€๊ฐ€ ์ž…๋ ฅ์œผ๋กœ ์‚ฌ์šฉ๋˜๋ฏ€๋กœ, ๊ธ€์ž ์˜ˆ์ธก ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์ด ๊ผญ์ง“์  ์˜ˆ์ธก ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์— ํฐ ์˜ํ–ฅ์„ ๋ฐ›์Œ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์—ˆ๋‹ค.

  • ๋‘ ๋ฒˆ์งธ ๋ฐฉ๋ฒ•

    1. ์‚ฌ์šฉ๋œ ์ด๋ฏธ์ง€ : ์›๋ณธ ์ด๋ฏธ์ง€์˜ ๋„ค ๊ผญ์ง“์  ์ขŒํ‘œ๋ฅผ x,y ๋ฐฉํ–ฅ์œผ๋กœ ๊ฐ๊ฐ ๋žœ๋คํ•˜๊ฒŒ ์ด๋™์‹œํ‚จ ํ›„ (128, 256)์˜ ํฌ๊ธฐ๋กœ ํˆฌ์˜๋ณ€ํ™˜ํ•œ ์ด๋ฏธ์ง€

    2. ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•: Salt & Pepper ๋…ธ์ด์ฆˆ, ๋ฐ๊ธฐ ์กฐ์ ˆ(์ „์ฒด ๋ฐ๊ฒŒ, ์ „์ฒด ์–ด๋‘ก๊ฒŒ, ๊ทธ๋ฆผ์ž) ์ˆ˜์ง‘๋œ ๋ฐ์ดํ„ฐ์…‹์€ ๋Œ€๋ถ€๋ถ„ ๋‚ฎ์— ์ฐ์€ ๋ฒˆํ˜ธํŒ ์ด๋ฏธ์ง€์˜€๊ธฐ ๋•Œ๋ฌธ์—, ํ…Œ์ŠคํŠธ ๋ฆฌํฌํŒ… ์‹œ ์•ผ๊ฐ„ ์ด๋ฏธ์ง€์— ๋Œ€ํ•ด์„œ๋Š” ์„ฑ๋Šฅ์ด ๋‚ฎ์•„์ง€๋Š” ๊ฒƒ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์—ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ฐ๊ธฐ ์กฐ์ ˆ ๋ฐ ๊ทธ๋ฆผ์ž ์ถ”๊ฐ€ ์ฆ๊ฐ• ๊ธฐ๋ฒ•์„ ์ถ”๊ฐ€ํ•˜์—ฌ ์—ฌ๋Ÿฌ ํ™˜๊ฒฝ์˜ ์ด๋ฏธ์ง€์— ๋Œ€ํ•ด ๊ฐ•๊ฑดํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋„๋ก ํ•˜์˜€๋‹ค.

    ์ถ”๋ก  ์‹œ ์‹ค์ œ ์ž…๋ ฅ๋˜๋Š” ์ด๋ฏธ์ง€ ์ฒซ ๋ฒˆ์งธ ๋ฐฉ๋ฒ• ๋‘ ๋ฒˆ์งธ ๋ฐฉ๋ฒ•

    ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์˜ ์˜ˆ์‹œ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค.

    ์‚ฌ์šฉ๋œ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ์ฆ๊ฐ•1 ๋ฐ์ดํ„ฐ์ฆ๊ฐ•2
  • Output : (๋ฐฐ์น˜์‚ฌ์ด์ฆˆ, 7, 45, 1) ๋ชจ์–‘์˜ ํ…์„œ
    7 -> 7๊ธ€์ž 45 -> 45๊ฐœ์˜ ๊ฐ€๋Šฅํ•œ ๋ฌธ์ž (['๊ฐ€', '๋‚˜', '๋‹ค', '๋ผ', '๋งˆ', '๊ฑฐ', '๋„ˆ', '๋”', '๋Ÿฌ', '๋จธ', '๋ฒ„', '์„œ', '์–ด', '์ €', '๊ณ ', '๋…ธ', '๋„', '๋กœ', '๋ชจ', '๋ณด', '์†Œ', '์˜ค', '์กฐ', '๊ตฌ', '๋ˆ„', '๋‘', '๋ฃจ', '๋ฌด', '๋ถ€', '์ˆ˜', '์šฐ', '์ฃผ', 'ํ—ˆ', 'ํ•˜', 'ํ˜ธ', '0', '1', '2', '3', '4', '5', '6', '7', '8', '9'])


5. pt >> onnx >> pb >> tflite ๋ณ€ํ™˜

  • YOLOv5
    ์ œ๊ณตํ•ด์ฃผ๋Š” export.py๋ฅผ ์‚ฌ์šฉํ•ด TensorFlow Lite ํŒŒ์ผ๋กœ ๋ณ€ํ™˜ํ•œ๋‹ค. ์ด ๋•Œ, Non Max Suppression ๋ถ€๋ถ„์€ TensorFlow Lite๋กœ ๋ณ€ํ™˜๋˜์ง€ ์•Š์•„ ์•ˆ๋“œ๋กœ์ด๋“œ ์ŠคํŠœ๋””์˜ค ์ฝ”๋“œ๋ฅผ ์งค ๋•Œ ๋”ฐ๋กœ ์ถ”๊ฐ€ํ•˜์˜€๋‹ค. YOLO์˜ ์ถœ๋ ฅ์œผ๋กœ ๋‚˜์˜ค๋Š” (1, 3024, 6)์˜ ํ…์„œ๋Š” 3024๊ฐœ์˜ ๊ฐ€๋Šฅํ•œ ๋ฐ”์šด๋”ฉ ๋ฐ•์Šค์™€, ๊ฐ ๋ฐ”์šด๋”ฉ ๋ฐ•์Šค์˜ x_center, y_center, width, height, confidence, ๊ฐ์ฒด ํด๋ž˜์Šค ์ •๋ณด๋ฅผ ํฌํ•จํ•˜๊ณ  ์žˆ๋‹ค. ์•„๋ž˜ ์ฝ”๋“œ๋Š” ๊ฐ€๋Šฅํ•œ 3024๊ฐœ์˜ ๋ฐ”์šด๋”ฉ ๋ฐ•์Šค ์ค‘ ๊ฐ€์žฅ ํฐ confidence ๊ฐ’์„ ๊ฐ€์ง€๋Š” ํ•˜๋‚˜์˜ ๋ฐ”์šด๋”ฉ ๋ฐ•์Šค๋งŒ์„ ์ถ”๋ก ์˜ ๊ฒฐ๊ณผ๋กœ ๋งŒ๋“œ๋Š” ์ฝ”๋“œ์ด๋‹ค (Non Max Suppression).
float max_conf = detectionResult[0][0][4];
        int idx = 0;
        for(int i = 0; i<3024; i++){
            if(max_conf < detectionResult[0][i][4]){
                max_conf = detectionResult[0][i][4];
                idx = i;
            }
        }
  • ๊ผญ์ง“์  ์˜ˆ์ธก ๋ชจ๋ธ & ๊ธ€์ž ์˜ˆ์ธก ๋ชจ๋ธ
    ๋ชจ๋ธ ํ•™์Šต ์‹œ, ๊ฒ€์ฆ ๋ฐ์ดํ„ฐ์…‹์— ๋Œ€ํ•ด ๊ฐ€์žฅ ๋†’์€ ์ •ํ™•๋„๋ฅผ ๊ฐ€์ง€๋Š” ๋ชจ๋ธ์˜ ๊ฐ€์ค‘์น˜๋ฅผ onnx ํŒŒ์ผ๋กœ ์ €์žฅํ•˜๊ณ , tflite_converter.py๋ฅผ ํ†ตํ•ด ์ตœ์ข…์ ์œผ๋กœ ์–ดํ”Œ๋ฆฌ์ผ€์ด์…˜ ์ƒ์—์„œ ๋ชจ๋ธ์„ ๋กœ๋“œํ•  ๋•Œ ์“ฐ์ด๋Š” TensorFlow Lite ํŒŒ์ผ๋กœ ๋ณ€ํ™˜ํ•œ๋‹ค.

6. ์•ˆ๋“œ๋กœ์ด๋“œ ์–ดํ”Œ๋ฆฌ์ผ€์ด์…˜ ์ œ์ž‘

์–ดํ”Œ๋ฆฌ์ผ€์ด์…˜์— ์•ž์„œ ๋งŒ๋“  ํ•™์Šต๋œ ๋ชจ๋ธ๋“ค์„ ์‚ฌ์šฉํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ๊ฐ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ถ”๋ก  ์ฝ”๋“œ๋ฅผ ๋งŒ๋“ค๊ณ , ์ด๋ฅผ ์•ˆ๋“œ๋กœ์ด๋“œ ์ŠคํŠœ๋””์˜ค์˜ MainActivity์— ๋ถˆ๋Ÿฌ์™€์„œ ์‚ฌ์šฉํ•ด์•ผ ํ•œ๋‹ค. ์šฐ๋ฆฌ๋Š” YOLOv5(DHDetectionModel.java), ๊ผญ์ง“์  ์˜ˆ์ธก(AlignmentModel.java), ๊ธ€์ž์˜ˆ์ธก(CharModel.java) ์ด ์„ธ ๊ฐ€์ง€ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ถ”๋ก  ์ฝ”๋“œ๋ฅผ ๋งŒ๋“ค์—ˆ๋‹ค. ์ถ”๋ก  ์ฝ”๋“œ์— ์‚ฌ์šฉ๋œ ๋ฉ”์†Œ๋“œ๋“ค์€ ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค:

  • ์ƒ์„ฑ์ž

    DHDetectionModel(Activity activity, Interpreter.Options options)
    AlignmentModel(Activity activity, Interpreter.Options options)
    CharModel(Activity activity, Interpreter.Options options)

    --> ๊ฐ ์ถ”๋ก  ์ธ์Šคํ„ด์Šค๋ฅผ ์ƒ์„ฑํ•  ๋•Œ, ๋ชจ๋ธ ์ธํ„ฐํ”„๋ฆฌํ„ฐ(mInterpreter)์™€ ๋ชจ๋ธ์— ๋“ค์–ด๊ฐ€๋Š” ์ž…๋ ฅ(mImageData)์— ๋Œ€ํ•ด์„œ ์ •์˜ํ•œ๋‹ค.

  • ๊ณตํ†ต์ ์œผ๋กœ ์‚ฌ์šฉ๋œ ๋ฉ”์†Œ๋“œ

    MappedByteBuffer loadModelFile(Activity activity)

    --> tflite ํŒŒ์ผ์„ ๋ถˆ๋Ÿฌ์˜ค๋Š” ๋ฉ”์†Œ๋“œ๋กœ ์ธํ„ฐํ”„๋ฆฌํ„ฐ ์ƒ์„ฑ์‹œ์— ์‚ฌ์šฉ๋œ๋‹ค.

    void convertBitmapToByteBuffer(Bitmap bitmap)

    --> ์ถ”๋ก ํ• ๋•Œ ์ด๋ฏธ์ง€๋ฅผ ๋ชจ๋ธ์— ๋“ค์–ด๊ฐ€๋Š” ์ž…๋ ฅ ํ˜•์‹์ธ ByteBuffer์˜ ํ˜•ํƒœ๋กœ ๋ฐ”๊พธ์–ด์ฃผ๋Š” ๋ฉ”์†Œ๋“œ์ด๋‹ค.

  • ์ถ”๋ก  ๋ฉ”์†Œ๋“œ

    • DHDetectionModel

      float[][] getProposal(Bitmap bm, Mat input)

      --> ์ด๋ฏธ์ง€๊ฐ€ ์ž…๋ ฅ์œผ๋กœ ๋“ค์–ด๊ฐ€๋ฉด float[2][5] ํ˜•ํƒœ์˜ ์ •๋ณด๋ฅผ ์ถœ๋ ฅํ•œ๋‹ค. ์ถœ๋ ฅ๊ฐ’์—๋Š” ๋ชจ๋ธ์ด ํƒ์ง€ํ•œ bounding box์˜ x, y, w, h, confidence์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ๋‹ด๊ณ  ์žˆ๋‹ค. Yolov5์— nms๊ฐ€ tflite ํ˜•ํƒœ๋กœ ๋ณ€ํ™˜๋˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์— ๋”ฐ๋กœ nms ์ฝ”๋“œ๋ฅผ ์ถ”๊ฐ€ํ•˜์˜€๋‹ค.

    • AlignmentModel

      float[] getCoordinate(Bitmap bitmap)

      --> DHDetectionModel์—์„œ ๋‚˜์˜จ ์ถœ๋ ฅ์„ ์ด์šฉํ•ด bounding box์˜ ํฌ๊ธฐ๋กœ ์ž๋ฅธ ์ด๋ฏธ์ง€๊ฐ€ ์ž…๋ ฅ์œผ๋กœ ๋“ค์–ด๊ฐ€๋ฉด, float[8] ํ˜•ํƒœ์˜ ์ •๋ณด๋ฅผ ์ถœ๋ ฅํ•œ๋‹ค. ์ถœ๋ ฅ๊ฐ’์—๋Š” ๋ชจ๋ธ์ด ์˜ˆ์ธกํ•œ ๊ผญ์ง“์ ์˜ ๋„ค ์ขŒํ‘œ์˜ (x,y)๊ฐ’์„ ๋‹ด๊ณ ์žˆ๋‹ค.

    • CharModel

      String getString(Bitmap bm)

      --> AlignmentModel์—์„œ ๋‚˜์˜จ ์ถœ๋ ฅ์„ ์ด์šฉํ•ด ๋ฒˆํ˜ธํŒ ํฌ๊ธฐ๋กœ ์ด๋ฏธ์ง€๋ฅผ ์ž๋ฅธ ํ›„ ์ „๋‹จ๋ณ€ํ™˜์„ ์ด์šฉํ•ด ์ •๋ฉด์œผ๋กœ ๊ณง๊ฒŒ ํŽธ ์ด๋ฏธ์ง€๊ฐ€ ์ž…๋ ฅ์œผ๋กœ ๋“ค์–ด๊ฐ€๋ฉด, String ํ˜•ํƒœ์˜ ์ •๋ณด๋ฅผ ์ถœ๋ ฅํ•œ๋‹ค. ์ถœ๋ ฅ๊ฐ’์—๋Š” ๋ชจ๋ธ์ด ์˜ˆ์ธกํ•œ ๋ฒˆํ˜ธํŒ์˜ ๊ธ€์ž ์ •๋ณด๋ฅผ ๋‹ด๊ณ ์žˆ๋‹ค.

  • ์ถ”๋ก  ์†๋„(FPS) ๋ฌธ์ œ ๊ฐœ์„ 
    ์ดˆ๊ธฐ์— ๋ชจ๋“  ๋ชจ๋ธ๋“ค์„ ์•ฑ์— ์ ์šฉํ•˜์˜€์„ ๋•Œ, ํ•œ ์ด๋ฏธ์ง€๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ์‹œ๊ฐ„์ด ๋„ˆ๋ฌด ์˜ค๋ž˜๊ฑธ๋ ค์„œ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๋ฐฉ๋ฒ•์œผ๋กœ ์‹ค์‹œ๊ฐ„ ์ถ”๋ก ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜์˜€๋‹ค.

    1. YOLO ์ž…๋ ฅ ์ด๋ฏธ์ง€ ํฌ๊ธฐ ๊ฐ์†Œ (640, 480) -> (256,192)
    2. GPU ๋Œ€๋ฆฌ์ž ์‚ฌ์šฉ
    3. ๋ฉ€ํ‹ฐ์Šค๋ ˆ๋”ฉ
  • ์ตœ์ข… ๋ชจ๋ธ๋ณ„ & ์ „์ฒด ์ถ”๋ก ์‹œ๊ฐ„

    ๋ชจ๋ธ ์ถ”๋ก ์‹œ๊ฐ„(millisecond)
    ๋ฒˆํ˜ธํŒ ํƒ์ง€ ๋ชจ๋ธ 45
    ๊ผญ์ง“์  ์˜ˆ์ธก ๋ชจ๋ธ 82
    ๊ธ€์ž ๋ชจ๋ธ 86
  • ์–ดํ”Œ๋ฆฌ์ผ€์ด์…˜ ์˜ˆ

    ์˜ˆ์‹œ1 ์˜ˆ์‹œ2
    ์˜ˆ์‹œ1 ์˜ˆ์‹œ2

7. Google Play์— ์–ดํ”Œ๋ฆฌ์ผ€์ด์…˜ ๋“ฑ๋ก

๋‹ค์šด๋กœ๋“œ:

์„ค์น˜ ์ „ ์„ค์น˜ ํ›„
์˜ˆ์‹œ ์˜ˆ์‹œ2
This repository contains an implementation of the Permutohedral Attention Module in Pytorch

Permutohedral_attention_module This repository contains an implementation of the Permutohedral Attention Module

Samuel JOUTARD 26 Nov 27, 2022
A TensorFlow implementation of DeepMind's WaveNet paper

A TensorFlow implementation of DeepMind's WaveNet paper This is a TensorFlow implementation of the WaveNet generative neural network architecture for

Igor Babuschkin 5.3k Dec 28, 2022
Turning SymPy expressions into JAX functions

sympy2jax Turn SymPy expressions into parametrized, differentiable, vectorizable, JAX functions. All SymPy floats become trainable input parameters. S

Miles Cranmer 38 Dec 11, 2022
Official code release for ICCV 2021 paper SNARF: Differentiable Forward Skinning for Animating Non-rigid Neural Implicit Shapes.

Official code release for ICCV 2021 paper SNARF: Differentiable Forward Skinning for Animating Non-rigid Neural Implicit Shapes.

235 Dec 26, 2022
[AAAI-2022] Official implementations of MCL: Mutual Contrastive Learning for Visual Representation Learning

Mutual Contrastive Learning for Visual Representation Learning This project provides source code for our Mutual Contrastive Learning for Visual Repres

winycg 48 Jan 02, 2023
The codebase for our paper "Generative Occupancy Fields for 3D Surface-Aware Image Synthesis" (NeurIPS 2021)

Generative Occupancy Fields for 3D Surface-Aware Image Synthesis (NeurIPS 2021) Project Page | Paper Xudong Xu, Xingang Pan, Dahua Lin and Bo Dai GOF

xuxudong 97 Nov 10, 2022
A PyTorch Implementation of PGL-SUM from "Combining Global and Local Attention with Positional Encoding for Video Summarization", Proc. IEEE ISM 2021

PGL-SUM: Combining Global and Local Attention with Positional Encoding for Video Summarization PyTorch Implementation of PGL-SUM From "PGL-SUM: Combin

Evlampios Apostolidis 35 Dec 22, 2022
List of all dependencies affected by node-ipc malicious commit

node-ipc-dependencies-list List of all dependencies affected by node-ipc malicious commit as of 17/3/2022 - 19/3/2022 (timestamp) Please improve upon

99 Oct 15, 2022
The authors' implementation of Unsupervised Adversarial Learning of 3D Human Pose from 2D Joint Locations

Unsupervised Adversarial Learning of 3D Human Pose from 2D Joint Locations This is the authors' implementation of Unsupervised Adversarial Learning of

Dwango Media Village 140 Dec 07, 2022
Resources related to EMNLP 2021 paper "FAME: Feature-Based Adversarial Meta-Embeddings for Robust Input Representations"

FAME: Feature-based Adversarial Meta-Embeddings This is the companion code for the experiments reported in the paper "FAME: Feature-Based Adversarial

Bosch Research 11 Nov 27, 2022
OptaPlanner wrappers for Python. Currently significantly slower than OptaPlanner in Java or Kotlin.

OptaPy is an AI constraint solver for Python to optimize the Vehicle Routing Problem, Employee Rostering, Maintenance Scheduling, Task Assignment, School Timetabling, Cloud Optimization, Conference S

OptaPy 211 Jan 02, 2023
Performant, differentiable reinforcement learning

deluca Performant, differentiable reinforcement learning Notes This is pre-alpha software and is undergoing a number of core changes. Updates to follo

Google 114 Dec 27, 2022
Unofficial Tensorflow-Keras implementation of Fastformer based on paper [Fastformer: Additive Attention Can Be All You Need](https://arxiv.org/abs/2108.09084).

Fastformer-Keras Unofficial Tensorflow-Keras implementation of Fastformer based on paper Fastformer: Additive Attention Can Be All You Need. Tensorflo

Yam Peleg 10 Jan 30, 2022
pcnaDeep integrates cutting-edge detection techniques with tracking and cell cycle resolving models.

pcnaDeep: a deep-learning based single-cell cycle profiler with PCNA signal Welcome! pcnaDeep integrates cutting-edge detection techniques with tracki

ChanLab 8 Oct 18, 2022
Code for the CVPR2021 paper "Patch-NetVLAD: Multi-Scale Fusion of Locally-Global Descriptors for Place Recognition"

Patch-NetVLAD: Multi-Scale Fusion of Locally-Global Descriptors for Place Recognition This repository contains code for the CVPR2021 paper "Patch-NetV

QVPR 368 Jan 06, 2023
Source code for "FastBERT: a Self-distilling BERT with Adaptive Inference Time".

FastBERT Source code for "FastBERT: a Self-distilling BERT with Adaptive Inference Time". Good News 2021/10/29 - Code: Code of FastPLM is released on

Weijie Liu 584 Jan 02, 2023
Proximal Backpropagation - a neural network training algorithm that takes implicit instead of explicit gradient steps

Proximal Backpropagation Proximal Backpropagation (ProxProp) is a neural network training algorithm that takes implicit instead of explicit gradient s

Thomas Frerix 40 Dec 17, 2022
Traditional deepdream with VQGAN+CLIP and optical flow. Ready to use in Google Colab

VQGAN-CLIP-Video cat.mp4 policeman.mp4 schoolboy.mp4 forsenBOG.mp4

23 Oct 26, 2022
Churn prediction

Churn-prediction Churn-prediction Data preprocessing:: Label encoder is used to normalize the categorical variable Data Transformation:: For each data

1 Sep 28, 2022
A deep neural networks for images using CNN algorithm.

Example-CNN-Project This is a simple project showing how to implement deep neural networks using CNN algorithm. The dataset is taken from this link: h

Mohammad Amin Dadgar 3 Sep 16, 2022