Distract Your Attention: Multi-Head Cross Attention Network for Facial Expression Recognition

Abstract

This paper presents a novel facial expression recognition network, called Distract your Attention Network (DAN). Our method is based on two key observations in biological visual perception. Firstly, multiple facial expression classes share inherently similar underlying facial appearance, and their differences could be subtle. Secondly, facial expressions simultaneously exhibit themselves through multiple facial regions, and for recognition, a holistic approach by encoding high-order interactions among local features is required. To address these issues, this work proposes DAN with three key components: Feature Clustering Network (FCN), Multi-head Attention Network (MAN), and Attention Fusion Network (AFN). Specifically, FCN extracts robust features by adopting a large-margin learning objective to maximize class separability. In addition, MAN instantiates a number of attention heads to simultaneously attend to multiple facial areas and build attention maps on these regions. Further, AFN distracts these attentions to multiple locations before fusing the feature maps to a comprehensive one. Extensive experiments on three public datasets (including AffectNet, RAF-DB, and SFEW 2.0) verified that the proposed method consistently achieves state-of-the-art facial expression recognition performance. The DAN code is publicly available.

Context Transformer and Adaptive Method with Visual Transformer for Robust Facial Expression Recognition

“…EAC [28] uses randomly removing some facial regions to enhance the model's learning of the non-noise part and thus enhances the model's performance. DAN [47] adopts a multi-task joint approach to capture both local and global feature information. Compared to these methods, our proposed method captures local and global features based on the use of adaptive learning to retain tokens that are most relevant to expression categorization, thus achieving the effect of removing noisy and occluded regions and enhancing computational efficiency.…”

A comparative study of deep convolutional neural network-based facial expression regression

“…Two key observations of biological visual perception are the basis of DAN. It is observed that multiple facial expression categories have essentially similar beginnings in appearance with very small disparities, while facial expressions express themselves through multiple facial feature regions, but in order to discriminate each facial expression, a coding method for high order interaction between local features is needed, and with these issues as the main focus, three components of the DAN are proposed: the feature clustering network (FCN) , Multiple Attention Network (MAN) and Attention Fusion Network (AFN) [6].…”

Dual-Branch Cross-Attention Network for Micro-Expression Recognition with Transformer Variants

“…This branch is then combined with the local-global information extracted by MobileViT, From the visualization results in Figure 11, both optical flow and MobileViT focus more on the regions around the eyes and the corners of the mouth, while Swin Transformer focuses on the spatial information between the eyebrows, the cheeks, and the chin areas. If the optical flow map is fed into MobileViT (experiment (b)), the optical flow method and MobileViT focus on similar regions, which will cause the network to duplicate the and generate unnecessary redundant information [41]. In contrast, in experiment (a), the optical flow map is sent to Swin Transformer, enabling this branch to focus on complementary ME regions and extract detailed features more comprehensively.…”