Research On Facial Expression Recognition Based On Residual Networ

Aiming at improving the accuracy of facial expression recognition,this paper deeply studies the facial expression recognition algorithm based on residual network.The key research contents of this paper include:(1)In view of the insufficient attention of traditional convolutional neural network to image feature channels,which further restricts the accuracy of facial expression recognition,based on multi-layer residual network,this paper introduces a channel attention structure senet(squeeze and exercise network)into the residual block of residual network,and proposes an improved residual network se-resnet19 integrating senet channel attention mechanism for facial expression classification.Senet strengthens the attention of important expression feature channels,suppresses the signals interfering with the feature channels,and realizes the recalibration of expression feature channels,so as to improve the ability of expression feature extraction and recognition.Se-resnet19 also adopts ELU activation function to replace the original relu activation function to solve the gradient disappearance problem caused by relu and further enhance the feature acquisition and classification ability of the network.Finally,the effectiveness of the improved residual network se-resnet19 is analyzed through experiments.(2)Aiming at the problems of insufficient ability to capture global expression features of senet and complex design of full connection layer,this paper improves senet by referring to the relevant knowledge of image style,and designs a style-based calibration module SCM(style-based calibration module).This module introduces style pooling to replace the single mean pooling of senet,strengthens the capture ability of global expression information,and uses a single full connection layer to replace the complex full connection layer of senet.SCM integrates the style transmission and channel attention mechanism,applies style pooling to extract image style information from each channel of the feature map,estimates the recalibration weight of each channel through channel independent style fusion,and brings the relative importance of a single image style into the feature map.Then,based on the residual network resnet19,this paper introduces SCM module into the residual module,and proposes an improved residual network scm-resnet19 combined with SCM channel attention mechanism for expression recognition task.Scm-resnet19 also uses ELU activation function to replace the original relu activation function to avoid the disappearance of gradient and further strengthen the learning ability of feature network.The final design experiment proves the superiority of the improved residual network scm-resnet19 in the task of facial expression recognition.(3)This paper designs a facial expression recognition system,and the system user interface is written by Python and pyqt.The system calls the trained improved residual network for expression recognition,which can be used in practical scenes.