Partitioned Region-focused and Graph Convolution Hybrid Network for Macro- and Micro-expression Spotting

Recently, the analysis of facial macro- and micro-expressions has attracted the attention of researchers. However, spotting micro-expressions in long video sequences is a significant challenge due to their subtle and transient nature, which makes it difficult for current methods to pinpoint the precise time intervals of these fleeting expressions accurately. To address this issue, we propose the hybrid lightweight, partitioned region-focused and graph convolution network \(\text{(PRF-GCN)}\). The proposed model incorporates a partitioned region-focused convolutional approach that aligns feature extraction with regions of interest based on human attention mechanisms, to capture fine-grained motion cues effectively. The extracted features are then transformed into graph structures to facilitate relational modeling and reasoning through GCNs. By predicting the probability that each frame belongs to a micro-expression interval, the proposed framework offers enhanced temporal localization capabilities. Experimental results obtained on the \(\text{CAS(ME)}^2\) and \(\text{SAMM-LV}\) datasets demonstrate strong performance, with F1-scores of 0.2124 and 0.1733, respectively. In addition, extensive ablation studies further confirm the effectiveness of the proposed \(\text{PRF-GCN}\).