Multi-Scale Local Feature Fusion Network for Facial Expression Recognition

To solve the problem that facial expression recognition (FER) system in actual application scenariosis always interfered by complex background which lead to low accuracy, we designed a multi-scale local feature fusion network (MSLFnet) to improve the performance of FER in actual application scenarios. Middle-level facial features map are extracted from the backbone, and the middle-level local feature is generated by a patch-level local attention module, the network can obtain richer facial expressions. Experiments is carried out on the FER datasets RAF-DB and FER+ to verify the efficacy of the network. Experimental results show that the accuracy of the proposed network on RAF-DB and FER+ is 2.5% and 1% higher than original ResNet-18, proving the effectiveness of MSLFnet.