DLANet: A lightweight dual-stream framework with fine-grained spatio-temporal attention for micro-expression recognition

Micro-expression recognition (MER) remains challenging in practical settings because subtle facial deformations are easily obscured by appearance variation and slight head motion, discriminative evidence is temporally sparse around apex-related moments, and stronger spatio-temporal modeling often conflicts with deployment efficiency under limited-data conditions. To address these issues, we propose a Dual Lightweight Attention-guided Network (DLANet), a compact two-stage dual-stream framework for deployment-oriented MER. In the first stage, the Micro-expression Perceptive Appearance-Motion Dual Network (MP-AMDNet) performs motion-aware feature acquisition by combining dense optical-flow priors, weak appearance-motion consistency, and entropy-regularized motion attention with adaptive key-frame selection to preserve sparse but informative temporal evidence. In the second stage, the Spatio-temporal Fine-grained Attention Network (ST-FANet) performs lightweight spatio-temporal refinement through factorized channel-aware pseudo-3D modeling, multi-scale dilated feature fusion, and motion-guided deformable refinement, thereby enhancing subtle local deformations without relying on heavy global attention or full 3D processing. Furthermore, a composite objective combining focal reweighting, auxiliary alignment regularization, and temporal sparsity regularization improves robustness under small-sample and class-imbalanced settings. Extensive experiments on CASME II, SAMM, SMIC, and MEGC 2019 under both Single-Dataset Evaluation (SDE) and Composite-Dataset Evaluation (CDE) show that DLANet achieves competitive or improved recognition performance, including 0.8650 UF1 and 0.8450 UAR on MEGC 2019, while remaining compact with 3.40M parameters and a transparent efficiency profile. These results indicate that DLANet provides a favorable accuracy-efficiency trade-off for MER in resource-constrained settings.