Enhancement-suppression driven lightweight fine-grained micro-expression recognition

Abstract

Micro-expressions are short-lived and authentic emotional expressions used in several fields such as deception detection, criminal analysis, and medical diagnosis. Although deep learning-based approaches have achieved outstanding performance in micro-expression recognition, the recognition performance of lightweight networks for terminal applications is still unsatisfactory. This is mainly because existing models either excessively focus on a single region or lack comprehensiveness in identifying various regions, resulting in insufficient extraction of fine-grained features. To address this problem, this paper proposes a lightweight micro-expression recognition framework –Lightweight Fine-Grained Network (LFGNet). The proposed network adopts EdgeNeXt as the backbone network to effectively combine local and global features, as a result, it greatly reduces the complexity of the model while capturing micro-expression actions. To further enhance the feature extraction ability of the model, the Enhancement-Suppression Module (ESM) is developed where the Feature Suppression Module(FSM) is used to force the model to extract other potential features at deeper layers. Finally, a multi-scale Feature Fusion Module (FFM) is proposed to weigh the fusion of the learned features at different granularity scales for improving the robustness of the model. Experimental results, obtained from four datasets, demonstrate that the proposed method outperforms already existing methods in terms of recognition accuracy and model complexity.