Development of a Lightweight Temporal Convolutional Network for Microexpression Recognition

Microexpression recognition (MER) remains challenging because facial movements are short, subtle, and usually require computationally expensive temporal models. This paper presents a lightweight depthwise-separable Temporal Convolutional Network (DS-TCN) for MER on the CASME II dataset. The model uses a dual-stream design that combines grayscale facial-frame features with optical-flow motion features after face cropping, Eulerian video magnification, resizing, and SSIM-based key-frame selection. Standard temporal convolutions are replaced with depthwise separable temporal convolutions to reduce parameter count and floating-point operations while preserving sequence learning. Three-fold cross-validation produced a validation accuracy of 72.6% and a macro-F1 score of 67.58%. Compared with the standard TCN baseline, the proposed DS-TCN reduced parameters from 1.57 M to 0.219 M, FLOPs from 10.33 G to 1.50 G, and CPU inference time from 580 ms to 210 ms. The results show that lightweight temporal modeling can provide competitive MER performance with substantially lower computational cost.