[GoogleML] Face Recognition

2023. 10. 9. 13:38ใ†ArtificialIntelligence/2023GoogleMLBootcamp

 

 

 

What is Face Recognition?

recognition -> ๋” ์–ด๋ ค์šด ํƒœ์Šคํฌ

 

 

 

One Shot Learning

DB์— ์ €์žฅ๋œ ์‚ฌ๋žŒ ์ค‘ ์žˆ๋Š”์ง€ or ์•„๋‹Œ์ง€ ์˜ˆ์ธกํ•˜๋Š” ๊ฒƒ 

only ํ•œ ์žฅ์˜ ์‚ฌ์ง„ ๋งŒ์œผ๋กœ ์‚ฌ๋žŒ์„ ๊ตฌ๋ถ„ํ•ด์•ผํ•จ

๊ฒŒ๋‹ค๊ฐ€ ์ƒˆ๋กœ์šด ํŒ€์›์ด ์ถ”๊ฐ€๋  ๊ฒฝ์šฐ ๋งค๋ฒˆ ์žฌํ•™์Šต? -> ๋น„ํšจ์œจ

 

 

 

๋”ฐ๋ผ์„œ ์œ ์‚ฌ๋„ ํ•จ์ˆ˜๋ฅผ ํ•™์Šตํ•˜๋„๋ก ํ•จ

DB์— ์žˆ๋Š” ํŠน์ • ์‚ฌ๋žŒ๊ณผ ์œ ์‚ฌํ•  ๊ฒฝ์šฐ ๋งค์šฐ ์ž‘์€ ๊ฐ’

์—†๊ฑฐ๋‚˜ ๋‹ค๋ฅธ์‚ฌ๋žŒ - ํฐ ๊ฐ’

 

 

 

Siamese Network

๋‘ ์ด๋ฏธ์ง€๊ฐ€ ๊ฐ™์€ ์‚ฌ๋žŒ์ธ์ง€ ์•„๋‹Œ์ง€ 

์ธก์ •ํ•˜๋Š” ๋ฉ”ํŠธ๋ฆญ์œผ๋กœ d๋ฅผ ํ•™์Šตํ•œ๋‹ค (norm ๊ฐœ๋…)

 

 

 

distance ํ•จ์ˆ˜์˜ concept

 

 

 

Triplet Loss

SVM์— ๋‚˜์˜ค๋Š” margin concept

 

 

 

positive d ์ด 0.5๊ณ   

margin ์ด 0.2 ๋ผ๋ฉด

negative d ์€ ์ ์–ด๋„ 0.7์€ ๋„˜์–ด์•ผ ํ•œ๋‹ค๋Š” ์˜๋ฏธ! 

 

 

 

๊ตฌ์ฒด์ ์ธ loss func

 

 

 

randomํ•˜๊ฒŒ A(๋Œ€์ƒ) P(positive) N(negative) ๊ณ ๋ฅด๋Š” ๊ฒƒ์€ ์‰ฌ์›€

์–ด๋–ป๊ฒŒ triplet์„ ์„ ์ •ํ•˜๋Š”์ง€ trainํ•˜๋Š” ๊ฒƒ์€ ์–ด๋ ต๋‹ค

 

 

 

๋™์ผํ•œ ์‚ฌ๋žŒ์ด๋ผ๋ฉด, d๊ฐ€ ์ž‘๊ณ , 

๋‹ค๋ฅธ ์‚ฌ๋žŒ์— ๋Œ€ํ•ด์„œ๋Š” distance๊ฐ€ ํฌ๋‹ค 

 

 

 

Face Verification and Binary Classification

x๊ฐ€ ์•„๋‹ˆ๋ผ ์นด์ด!

 

 

 

๋‘ ๋„คํŠธ์›Œํฌ๋ฅผ ํ•™์Šต์‹œ์ผœ์„œ 

์ƒˆ๋กœ์šด ์‚ฌ๋žŒ์ด database์— ๋“ค์–ด์˜ค๋”๋ผ๋„ 

1. ์ „์ฒด ์ด๋ฏธ์ง€๋ฅผ ์ €์žฅํ•  ํ•„์š” X

2. precompute concept๋กœ, ๋งค๋ฒˆ ํ•™์Šตํ•˜๋Š” ๋น„ํšจ์œจ X

 

 

 

 

 

 

 

 

'ArtificialIntelligence > 2023GoogleMLBootcamp' ์นดํ…Œ๊ณ ๋ฆฌ์˜ ๋‹ค๋ฅธ ๊ธ€

[GoogleML] Convolutional Neural Networks ์ˆ˜๋ฃŒ  (0) 2023.10.11
[GoogleML] Neural Style Transfer  (2) 2023.10.11
[GoogleML] Object Detection  (1) 2023.10.09
[GoogleML] Convolutional Neural Networks Case Studies  (1) 2023.10.06
[GoogleML] Convolutional Neural Networks  (1) 2023.10.04