Facial Expression Recognition Based on Improved VGG16 Convolutional Neural Network

Abstract

Facial expression recognition has the characteristics of complexity and variability, which makes it a very popular research topic at present. Based on the VGG16 convolutional neural network, a new deep learning expression recognition method is proposed, which can obviously improve the disadvantage of low accuracy of traditional expression recognition methods. The new network is based on the basic structure of the VGG16 network, meanwhile uses a single graphics processing unit for training. Firstly, the VGG16 network is divided into 5 Blocks, and then the last 3 Blocks are fused with features, and the Spatial Group Enhance (SEG) attention module is added. Finally, the redundant fully connected layers were deleted, and the final classification result was output by one fully connected layer, which effectively reduced the parameters of the neural network. Experimental results on FER2013 dataset and CK+ dataset show that the recognition rate of the new network for facial expression reaches 68.85%and 97.46%respectively, which is higher than that of other traditional facial expression recognition methods.