---
tags: 生物辨識
---

# Frequency-aware Discriminative Feature Learning Supervised by Single-Center Loss for Face Forgery Detection
握靠這名字也太長

## Framework

## Adaptive frequency features generation module (AFFGM)
- `RGB` images are transformed into `YCbCr` color space.

> 人眼細胞的感知能力分為`色彩`及`亮度`
> 而人眼對`亮度`的敏感程度高為`色彩`的敏感程度
> `RGB` space 容易受到光源變動的影響，所以採用 `YCbCr` 將亮度切割

- `2D DCT` transformation
將空間轉換到`頻率`上，將高頻率空間進行壓縮及簡化

[DCT相關教學視頻](https://www.bilibili.com/video/BV17M4y1u7Ek?from=search&seid=3264094561220943602)

## Adaptive frequency information mining block (AFIMB)

- 三個 Conv
- 一個 Max pooling
- 兩個 linear layer

## Single-center loss (SCL)
$$
L_{sc} = M_{nat}+max(M_{nat}-M_{man}+m\sqrt D, 0)
$$

- $M_{nat}$：Natural Face 到 Center 的 ED.
- $M_{man}$：Manipulated Face 到 Center 的 ED.

$$
M_{nat} = \frac{1}{|\Omega_{nat}|}\sum_{i\in\Omega_{nat}}\|f_i-C\|_2 \\
M_{man} = \frac{1}{|\Omega_{man}|}\sum_{i\in\Omega_{man}}\|f_i-C\|_2
$$
- $C$：Center point of natural faces

### Total loss
$$
L_{total}=L_{softmax}+\lambda L_{sc}
$$


## 筆記