---
tags: 生物辨識
---
# Suppressing Uncertainties for Large-Scale Facial Expression Recognition

## Problem to solve
- Annotating a qualitative large-scale facial expression dataset
- ambiguous facial expressions
- low-quality facial images
- the subjectiveness of annotators

## Contribution
- 提出了一個Self-Cure Network(SCN)去減少不確定性的影響
- 設計了一個self-attention的機制使得SCN可以學到重要和有意義的feature
- 並利用self-attention的機制重新標注uncertained data


## Pipeline of Self-Cure Network
- Extract the deep features by a backbone network, let $F$ denotes the facial features of $N$ images.
$$
F = [x_1,x_2,...,x_N]\in \mathbb{R}^{N\times D}
$$
- Self-Attention Importance Weighting
- Let $\alpha_{i}$ denote the importance weight of the i-th sample, $W_\alpha$ is the parameters of the FC layer used for attention, $\sigma$ is the sigmoid function.
$$
\alpha_{i}=\sigma(W_\alpha^Tx_{i})
$$
- Logit-Weighted Cross-Entropy Loss
$$
L_{WCE} = - \frac{1}{N} \sum_{i=1}^{N}log(\frac{e^{\alpha_{i}W_{y_{i}}^Tx_{i}}}{\sum_{j=1}^{C}e^{\alpha_{i}W_{j}^Tx_{i}}})
$$
- Rank Regularization
- rank the learned attention weights in descending order and then split them into two groups with a ratio β, $M=βN$
$$
L_{RR} = max(0, \delta_1-(\alpha_{H}-\alpha_{L})), \alpha_{H}=\frac{1}{M}\sum_{i=1}^{M}\alpha_{i},\alpha_{L}=\frac{1}{N-M}\sum_{i=M}^{N}\alpha_{i}
$$
- total loss
$$
L_{total} = \gamma L_{WCE} + (1-\gamma)L_{RR}
$$
- Relabeling Mechanism
$$
y_{new} =
\begin{cases}
l_{max}, & \mbox{if}P_{max}-P_{gtInd}>\delta_{2} \\
l_{org}, & \mbox{otherwise}
\end{cases}
$$

## Experiments
- Visualization of the learned importance weights in SCN