Lightweight-efficient facial expression recognition with MSDE strategy for complex illumination

This paper presents a lightweight and efficient method designed for facial expression recognition (LEFER) under complex lighting conditions. Existing lightweight facial expression recognition methods have poor robustness when dealing with environmental changes, whereas methods with illumination invariance often face heavy computational burdens. LEFER integrates the multi-stage data enhancement (MSDE) strategy and the lightweight backbone network (LEFER-Net), which addresses this gap. The MSDE strategy combines geometric enhancement, photometric enhancement based on local Gaussian shadows, adaptive normalization, and the Mixup strategy to dynamically standardize input data, effectively addressing interferences such as spatial deformation, illumination fluctuations, and contrast attenuation. LEFER-Net involves three feature refinement stages and a classification head. First, the channel attention module based on local cross-channel interaction (LCCISCA) efficiently calibrates features. Second, the Gaussian mixture spatial attention (GMSA) module introduces a parameterized two-dimensional Gaussian kernel matrix to focus on key facial regions. Experimental results on FER-2013 and RAF-DB verify that LEFER achieves higher recognition accuracy while maintaining low parameter count and computational complexity. Ablation experiments and visualization results confirm that the proposed components effectively improve recognition accuracy under complex lighting conditions, achieving a favorable balance among recognition performance, illumination adaptability, and deployability.