--- tags: 生物辨識 --- # Learning Generalized Spoof Cues for Face Anti-spoofing ## Introduction github: https://github.com/VIS-VAR/LGSC-for-FAS 百度在2020提出的論文,號稱在熱門的RGB資料集取得SOTA #### 主要思路為以下兩點 1. Anomaly detection 把Anti-spoofing視為異常偵測,希望活體可以在live center c越靠近,而非活體遠離 ![](https://i.imgur.com/sjNFJjZ.png) 2. Residual learning 把區分活體的feature當成殘差(spoof cues) 並且spoof cues只存在spoof sample H(x) = F(x) + x ## Method ![](https://i.imgur.com/1Dy0ieK.png) Triplet loss: Anchor只會是正樣本,直接各層取global average pooling,對應等式4 ![](https://i.imgur.com/pLgHjAE.png) Regression loss: 只拿正樣本產生出來的spoof cue跟zero map算pixel-wise L1 loss,對應等式3 ![](https://i.imgur.com/N7AdKqq.png) Classification Loss 加了這個可以當spoof cue amplifier ![](https://i.imgur.com/dQwhF55.png) ![](https://i.imgur.com/jgcnpvZ.png) ## Experiments Inter class testing ![](https://i.imgur.com/7hZVfZp.png)