Facial Expression Recognition Based on Improved Convolutional Neural Network

Abstract

The accuracy of traditional convolutional neural network (CNN) for facial expression recognition (FER) is not high. To improve detection accuracy, a face micro-expression recognition algorithm combining image deflection angle weighting, which achieves static and dynamic facial expression recognition, was proposed. First, face recognition was performed on the face to be measured based on the Haar features in the OpenCv library. Second, pre-processing, such as face position detection, face cropping, normalization, and data enhancement, was performed on the measured image to avoid irrelevant information interfering with the judgment. Third, convolutional neural network was used for FER, and the result of the linear weighting of expression labels measured by deflecting the face to be tested by multiple angles was used as final recognition result to improve the accuracy. Lastly, a camera was used for real-time judgments and static recognition on the CK+ data set. Results show that classifying difficult images in multiple combinations and building integrated models improve prediction accuracy. The recognition rate on the CK+ data set is 97.85%, which is an improvement of about 3% compared with the cross-connect LeNet-5 network algorithm, thereby verifying its feasibility and effectiveness. This study provides a good reference for the improvement of facial expression detection performance.