Abstract

This paper proposes a unique pipeline for micro-expression recognition using a Dual-Path 3D Convolutional Neural Network enhanced with Hybrid Attention and Squeeze-and-Excitation Blocks. The three main goals of the pipeline are to (1) Optimize the extraction of spatial-temporal features using advanced neural network architectures, (2) Enhance data representation by implementing targeted image augmentation and balanced class distribution, and (3) Enhance feature fusion using state-of-the-art network techniques. Comprehensive experiments were conducted on four benchmark datasets: CAS(ME)2, SMIC, SAMM, and CASME II. The Hybrid Attention-3DNet model demonstrated superior recognition accuracy of 93.95% for CAS(ME)2, 93.42% for SMIC, 93.61% for SAMM, and 93.79% for CASME II, surpassing the state-of-the-art methods across these datasets. These outcomes demonstrate the efficacy and robustness of the proposed pipeline, underscoring its potential for a range of micro-expression recognition uses.

Key words: Micro-Expression Recognition, 3D Convolutional Dual Path Network, Hybrid Attention, Squeeze-and-Excitation Blocks, Deep Learning.