Enhancing Facial Micro-Expression Recognition in Low-Light Conditions Using Attention-guided Deep Learning

This paper presents a novel approach to facial micro-expression recognition using an attention-guided deep learning framework in low-light conditions. Micro-expressions, characterized by their subtle and rapid nature, provide valuable insights into genuine emotions but are challenging to detect, especially in suboptimal lighting. We propose the Attention-Guided Micro-Expression Network (AGMENet), which integrates a Low-Light Enhancement Module (LLEM) with a modified DenseNet architecture and multi-head attention mechanism. The LLEM effectively improves image quality while preserving crucial facial movements, enabling robust feature extraction in low-light scenarios. Our attention mechanism focuses on the most relevant facial regions and temporal segments, significantly enhancing recognition accuracy. We introduce a temporal ensemble method to leverage information from multiple frames within micro-expression sequences, addressing challenges associated with their brief duration. Extensive experiments on three benchmark datasets (CASME II, SAMM, and SMIC) demonstrate the superiority of AGMENet over state-of-the-art methods, achieving an average accuracy improvement of 4.38% in low-light conditions. Ablation studies validate the effectiveness of each component, while attention map visualizations provide insights into the model's decision-making process. The proposed approach shows promise for real-world applications in security, healthcare, and human-computer interaction, where accurate emotion recognition in varying lighting conditions is crucial.