SDNET: LIGHTWEIGHT FACIAL EXPRESSION RECOGNITION FOR SAMPLE DISEQUILIBRIUM

Facial expression recognition (FER) based on the convolutional neural network (CNN) in the wild have numerous challenges. For instance, the complexity of the network model makes FER tasks difficult to deploy on portable devices. Some approaches design lightweight networks to reduce the model size, while the intrinsic imbalance of the existing facial emotion datasets is still ignored. In order to overcome the above problems, the lightweight CNN based on sample equalization method for FER is designed to reduce the network parameters sharply while maintaining the identification accuracy. Specifically, to reduce the number of network parameters, a lightweight network framework (SDNet) is designed with separable convolution layers and dense blocks, which can significantly reduce network parameters. Second, the adaptive class weights are proposed to solve the imbalance of sample numbers. Moreover, a resist overfitting (RO) loss function is proposed to improve the classification accuracy. Extensive experiments are conducted on lab-controlled datasets (CK+, Oulu-CASIA) and in-the-wild datasets (FER2013, SFEW). Experimental results show that our method is superior to several state-of-the-art FER methods.