DIR-EC: intra-region enhancement and inter-region collaboration network for facial expression recognition

Juan Liu, Ying Wang, Zhong Huang, Min Hu, Fuji Ren, Wenquan Xu

Abstract: To suppress the influence of occlusion and pose variation on facial expression recognition (FER) in natural scenes, a FER network based on intra-region enhancement and inter-region collaboration (DIR-EC) is proposed. The proposed network mainly contains intra-regional multi-scale enhancement subnet (Intra-MSEnet), inter-regional multi-granularity collaborative subnet (Inter-MGCnet), and adaptive fusion subnet (AFSnet). In the Intra-MSEnet, a down-up dual attention mechanism is constructed to extract the intra-regional multi-scale features with low-level spatial semantics from up to down and high-level channel semantics from down to up. In the Inter-MGCnet, a collaborative guidance attention structure is designed to capture inter-regional multi-granularity collaborative semantics of local features and global features, which achieves the guidance of coarse-grained features to fine-grained features. In the AFSnet, an adaptive fusion strategy is proposed to fuse inter-regional collaborative semantics and global guidance semantics. The experimental results show that the expression recognition rates of DIR-EC are 90.14% and 90.32% on RAF-DB and FERPlus datasets, which are 13.71% and 11.01% higher than the baseline method, respectively. Compared with related methods, the proposed DIR-EC improves the expression recognition performance in natural scenes, and reduces the influence of occlusion and pose variation. The code will be available at https://github.com/liujuanjuanliu/DIR-ECnet.

External IDs:dblp:journals/vc/LiuWHHRX25