Micro-Expression Recognition Network Based on Attention Mechanism and Dual-Channel Fusion

Abstract

: Micro-expression recognition plays a significant role in psychological analysis and public safety. To address the challenges of short duration and subtle motion interference in micro-expressions, this paper proposes an Asymmetric Dual-channel Fusion Network (ADF-Net) based on attention mechanisms. The method employs the Convolutional Block Attention Module (CBAM) to adaptively enhance the spatial structural features of static images, integrates Spatial Pyramid Pooling (SPP) to capture multi-scale motion details in dynamic images, and utilizes Squeeze-and-Excitation (SE) attention to achieve adaptive recalibration of dual-channel weights. Experimental results demonstrate that the proposed method achieves UF1 scores of 0.9062, 0.7718, and 0.7099 on the CASME II, SAMM, and SMIC benchmark datasets, with a composite UF1 score of 0.7513 on the 3DB-combined dataset.