Research On Facial Expression Recognition Method Based On Deep Learning And Confidence

The research on facial expression recognition has always been concerned by research and industry,but the traditional manual feature methods have some problems such as cumbersome steps,low efficiency and so on.The efficient network model based on cascade framework and attention mechanism is proposed,combined with a confidence-based loss function algorithm to recognize and classify facial expressions.The main work of the thesis is as follows:(1)Aiming at the low accuracy rate of facial expression recognition and difficulty in extracting effective features,the method of facial expression recognition combining cascade framework and attention mechanism is proposed.Firstly,the cascade frame model is used to eliminate the redundant interference of the face image;secondly,the attention mechanism is embedded in the deep effective network to infer the attention weight from the channel and spatial dimension,so as to enhance the expressiveness of expression features and suppress the influence of redundant information;finally,the focus loss function is used to further weaken the impact of the uneven distribution of data sets.Based on the data set FER2013 and JAFFE,the proposed method has achieved 72.42%and 95.71%recognition accuracy respectively,which is progressiveness and superior compared with other algorithms.The experimental results show that the proposed model has a better effect on extracting facial expression recognition features.(2)Aiming at the problem of unbalanced distribution of sample data in the facial expression dataset,two staged loss function fine-tuning algorithms are proposed from the perspective of confidence.The phased fine-tuning algorithm based on the focus loss function first screens the mislabeled samples to reduce the impact on the model,and then solves the problem of insufficient model learning caused by the confidence through the phased training strategy,thus improving the model classification performance.In addition,the staged loss fine-tuning algorithm based on gradient coordination mechanism suppresses the contribution of easily separated samples and abnormal samples to the model through gradient density,and focuses more on the contribution of difficultly separated samples,thus improving the model recognition rate.Based on the data set FER2013 and JAFFE,the recognition accuracy of the two methods has been improved by 0.44%,0.86%and 1.02%,1.43%respectively compared with the focus loss function method.The experimental results show that the two staged loss function fine-tuning algorithms proposed from the perspective of confidence are effective in dealing with the problem of uneven sample distribution and difficult sample separation.(3)A facial expression recognition system is designed and implemented.The facial expression recognition system designs the frontend by using the Vue framework,and builds the back-end system based on the Django framework.And the system uses the method based on the cascade framework and attention mechanism for facial expression recognition,with the loss function algorithm based on confidence as the recognition model.