Research On Expression Recognition Algorithm Based On Deep Learning

In recent years,the facial expression recognition technology has been widely used in many fields,involving every aspect of our daily life,such as can be judged in the field of commercial customers liking of goods,in the field of medical rehabilitation can help better rehabilitation,patients with mental illness can be monitored in the field of education of the students' concentration in class,in the field of intelligence traffic can prevent fatigue so as to reduce the number of traffic accidents.However,in the process of practical application,the collected facial expression images have different light intensity and too many complex backgrounds,which leads to the reduction of facial expression recognition rate and is difficult to meet the actual demand.In order to solve this problem,this paper proposes an improved convolutional neural network based on Le Net convolutional neural network(CNN)and the idea of continuous convolution,and further enhances the robustness of facial expression recognition through preprocessing such as face detection,clipping and normalization.The following are the main research contents of this paper:1.For the problem of low expression recognition rate,this paper improves the classical Le Net convolutional neural network by using the idea of continuous convolution,and thus proposes a new convolutional neural network(Succession Net model)with continuous convolution.Continuous convolution can enhance the nonlinear expression ability of convolutional neural network,so as to improve the robustness of expression recognition.In order to verify the advantage of Succession Net model,the algorithm model is compared with Le Net model and discontinuous convolutional neural network(Deep Net model).At the same time,this algorithm model is compared with other classical convolutional neural network models.Finally,experimental results show that the recognition rate of the proposed algorithm is improved compared with Le Net and Deep Net.Although the training convergence speed is not as fast as the original shallow network Le Net,it is significantly improved compared with Deep Net.By comparing with other classicalnetworks,it can be seen that the recognition rate of this algorithm has some advantages.2.In view of the problems of different light intensity and too many complex backgrounds in facial expression images collected in the process of practical application,this paper adopts the Face r-cnn method in the deep learning method for Face detection,which is based on the Faster r-cnn framework and optimized for the particularity of Face detection.By setting the size range of the face detection box,the detected face is clipped into two sets of images of different sizes and saved to the folder.Then the two groups of images were normalized and preprocessed.After that,the two groups of images containing complex background and only face information were respectively input into the Succession Net model proposed in this paper for training.Finally,the effectiveness of the image preprocessing method in this paper could be verified by the training recognition rate curve.