Facial Expression Recognition Methods Based On ResNet

Facial expressions are an essential way for humans to express emotions and communicate with each other.Automatic facial expression recognition is widely used in the fields of human-computer interaction including intelligent driving,assisted medical care,and intelligent control.In recent years,with the development of deep learning technology,facial expression recognition research has made great progress.However,the accuracy of facial expression recognition used in the wild still needs to be further improved duo to some problems such as occlusion,posture changes and background interference and so on.In this paper,using ResNet as the basic architecture,we propose the improved facial expression methods by introducing some optimization measures like joint optimization,attention mechanism and second-order pooling layers.The extensive experiments have been conducted on the public datasets,indicating the high accuracy and robustness of the proposed methods.Additionally,the paper also operates the design and realization of the interactive website of facial expression recognition.The main contents are as follows:(1)Based on the ResNet,the joint optimization strategies are proposed.The joint optimization strategies are the filter response normalization and batch normalization,instance normalization and group normalization,group normalization and batch normalization embedded in the network,respectively,to make up for the shortcomings of single normalization.That is as far as possible to retain effective feature information,to balance and improve the distribution of feature data,and enhance model performance.The model achieves the accuracy rates of 73.558% and 94.9% on the FER2013 and CK+,respectively,which verifies the effectiveness of the proposed method.(2)Based on the ResNet and joint normalization,the Convolution Module Attention Mechanism(CBAM)is introduced to extract features with strong expressive ability in the channel and space,and reduce the weight of irrelevant features,thereby improving the model's performance.The accuracy rates of73.753% and 95.7% are achieved on the FER2013 and CK+,respectively,showing good competitiveness.(3)Based on the ResNet,we combine the Bottleneck Attention Mechanism(BAM)and the global second-order pooling layer into the network.The BAM enables network to focus on the extraction of important features of expression.The global second-order pooling layer is able to make full use of the second-order statistic information to explore the correlation between expression features.On this basis,joint normalization is used to balance and improve the distribution of feature data and the accuracy of expression recognition.The proposed method achieves the accuracy rates of 74.227%,95.8% and 56.1%,respectively,and the performance is better than many existing mainstream methods.The results demonstrate that the proposed model has good accuracy and robustness.