Enhanced Feature Extraction and Attention Mechanisms for Micro-Expression Recognition

Micro-expressions refer to subtle changes in expressions that are displayed by humans in a very short period of time. As a form of non-verbal emotional expression, micro-expressions can more accurately reflect an individual's true inner feelings. We propose a shallow dual-branch 3D-CNN architecture for the first stage of feature extraction. Additionally, we enhance and optimize the channel attention module within the convolutional block attention module. We then employ Gated Recurrent Unit and Multi-Scale Multi-Head Self-Attention for the second stage of feature extraction. We tested our model on 3 different micro-expression databases and obtained competitive results. Numerous experimental results show that our method can achieve good performance with relatively simple input and architecture. The source code and pretrained models are available at https://github.com/dannyFan-0201/MLSP_2024.