Composite loss collaborative optimization for compound facial expression recognition
Nihan Guo · 2025
To address the core challenges of intra-class feature dispersion and inter-class semantic ambiguity in compound facial expression recognition under natural scenes, this paper proposes a multi-loss collaborative optimization framework that integrates Adaptive Enhanced Center Loss (AECL) and Confusion-Guided Island Loss (CGIL). AECL enhances intra-class compactness through dynamic hard sample mining and class-balanced strategies, while CGIL improves inter-class separability via dual-metric constraints and a high-frequency misclassification penalty mechanism. Combined with cross-entropy loss, the proposed framework forms a three-level optimization system of “classification guidance – intra-class aggregation – inter-class separation.” Experiments demonstrate that the proposed method achieves 61.23% accuracy on the RAF-DB compound subset, outperforming the current state-of-the-art by 0.9%. At the meantime, our method achieves the accuracy of 87.04% for the basic expression recognition task, providing an efficient solution for fine-grained affective computing in fields such as intelligent healthcare and human-computer interaction.