| Expression Recognition refers to the ability of machines to judge human emotions through facial expressions.It is very important for the development of AI.It is applied in the fields of user preference analysis,assisted driving and case detection.At present,Expression Recognition can perfectly recognize the facial expressions collected in the laboratory.However,for the real environment,the recognition accuracy decreases sharply due to the influence of complex background information,light,posture and occlusion.This paper uses the phased processing method of detection and recognition multi network to design and manufacture the Expression Recognition System applied in the real environment.Firstly,promote the diversity of data sets through collect and search data,and expansion processing.Then build the Convolutional Neural Network by using Python language and Torch deep learning framework.The face detection network is designed by using the excellent performance YOLOv4 algorithm.The network is trained on the Wider Face dataset.The experimental results show that the AP50 is 87.4% and the FPS is 20.5s.It can achieve the best balance between accuracy and detection speed.Then,a feature extraction classification network based on RESNET is proposed,which adds three attention modules and multi-scale fusion structure.The network training results show that the recognition accuracy of the network is significantly improved compared with the original network,reaching 95.65% and85.19% on the occluded CK+ and RAF-DB dataset,with an increase of 5.17% and 5.73%respectively.Finally,a system interface is designed by using wx Pthon toolkit,and the expression recognition system is made by using the network model.The system is tested by taking photos in the real environment.The recognition accuracy of the system is stable at more than 78%,which can meet the requirements of practical application.The expression feature extraction and classification network designed in this paper has high recognition accuracy and performance advantages.The expression recognition system built by using this network is easy to operate and can overcome the influence of certain face pose changes and occlusions.It has practical application value. |
