Studies On Application Of Deep Learning Combined With Support Vector Machine In Facial Expression Recognition

As an important part of intelligent human-machine interaction, facial expression gradually is used in the hot areas such as computer vision, artificial intelligence, and researchers in different fields put forward many new methods to improve its recognition accuracy. But usually the facial expression image must be face, limiting its scope of application. In the research of machine learning, as a new field, deep learning processes data chiefly through feature of unsupervised automatic extraction.To the low accuracy and limitations of facial expression recognition of traditional research, in this paper, an improved multi-layer classification and explaining model based on deep learning and support vector machine was proposed. First of all, the model was trained to build a facial expression recognition model(RBM-SVM model) by training set of facial expression sample selected from the CMU Face Images each expression contains 8 different angles(Presence of wearing sunglasses expression); then verifying the performance of the test sample in different situations: compare the recognition results of RBM-SVM, DBN and SVM over different image resolution, hidden layers number and the number of nodes, and studying the relationship between RBM-SVM and support vectors or the accuracy under different hidden layers number, the number of nodes, The influence of parameter penalty factor C and kernel function parameters on the recognition results of support vector machine is analyzed.Under the same condition, the facial expression recognition by RBM-SVM has a higher accuracy than by DBN or by pure SVM; RBM-SVM is not sensitive for image resolution, that is, there is a tiny difference in facial expression identification result for different resolutions. The higher the number of hidden layers, the higher the recognition accuracy is. With the improvement of the penalty factor C, the recognition accuracy of the algorithm is the highest and no longer changes when the proper value is obtained; the accuracy of facial expression recognition for test samples is the highest 96.4% when the kernel function parameters is 0.004.