High dynamic range preprocessing, ParallelAttention Transformer and CoExpression analysis for facial expression recognition

Abstract

Facial expression recognition (FER) aims to enable computers to automatically detect and recognize human facial expressions, thereby understanding their emotional states. Despite significant technological advancements in recent years, FER tasks still face several challenges, including expression diversity, individual differences, and the impact of lighting and detail variations on recognition accuracy. To address these challenges, a high-performance FER model is proposed that comprises three key components: High Dynamic Range (HDR) Preprocessing Module, ParallelAttention VisionTransformer structure, and CoExpression Head. In the preprocessing stage, the HDR Preprocessing Module optimizes input images through local contrast and detail enhancement techniques, improving the model’s adaptability to lighting and detail variations. During the feature processing stage, the ParallelAttention VisionTransformer structure employs a multi-head self-attention mechanism encoder to effectively capture and process facial expression features at various scales, allowing for a detailed understanding of subtle facial expression differences. Finally, the CoExpression Head utilizes a collaborative expression mechanism to efficiently handle and refine features across different expression states during the feature integration process. Combining these three stages significantly enhances the accuracy of facial expression recognition. Extensive experimental evaluations on public datasets, RAF-DB and AffectNet, demonstrate that the model achieves accuracy rates of 92.11%, 67.25%, and 63.40% on RAF-DB, AffectNet, and AffectNet-8, respectively, exhibiting outstanding performance comparable to other state-of-the-art models.