Research On Facial Expression Recognition Based On Deep Learning

With the spiral rise of intelligent era and deep learning technology,facial expression recognition technology has become one of the important research topics in the field of computer vision.It is widely used in safe driving,auxiliary medical treatment,criminal interrogation,video recommendation and other fields.While traditional methods for facial expression recognition have some limitations,convolutional neural network(CNN)have strong ability to extract facial expression features autonomously.Therefore,this paper mainly studies the facial expression recognition algorithm based on convolutional neural network.In view of the problems existing in the convolutional neural network processing facial expression classification task,the following two aspects are mainly done:(1)Aiming at the problems of complex model structure,excessive training parameters and low recognition accuracy when the current convolutional neural network uses its autonomous learning ability to process facial expression recognition tasks,a facial expression recognition algorithm based on the combination of improved CNN and support vector machine(SVM)is proposed.Firstly,the network model was built based on the design concept of VGGNET-16 continuous convolution.In order to obtain more nonlinear features,the series and parallel fusion of small-size convolution kernels was adopted to replace the action of large-size convolution kernels.Secondly,in order to solve the problem of too many parameters generated by the full connection layer in traditional CNN,the adaptive global average pooling(GAP)layer is used instead.Finally,using the advantages of SVM in small sample and multi-classification tasks,the facial expression classification task is completed by using SVM classifier instead of softmax function in traditional CNN.Experimental results show that the recognition accuracy of the established network model is 73.4% and 98.06% on FER2013 and CK+ data sets respectively,which is nearly 1.2 times higher than the traditional CNN model.Moreover,the network model has simple structure,ideal recognition effect and good robustness.(2)Considering that the key areas of facial expression recognition are eyes,eyebrows,mouth,etc,and the convenience of attention mechanism plug and play can focus on the key information in all feature images,a facial expression recognition algorithm is proposed to introduce attention mechanism and feature fusion.Firstly,a lightweight convolution structure is constructed to reduce the number of parameters and the influence of overfitting.Secondly,in view of the different importance of the image information contained in each feature channel and the spatial difference between the internal pixels of feature elements,on the one hand,the embedded channel attention branch CA strengthens the features of key channels and weakens the low-frequency channel information.On the other hand,the spatial attention module SA is serial connected to adaptively capture the spatial attention of core pixels in different feature images,so as to more accurately locate key expression regions and improve the processing efficiency of the model.Finally,The open data set RAF-DB was tested on the CASACNN model after feature fusion.The results show that the recognition accuracy is improved by 2.76 percentage points,and it has strong classification performance and robustness.There are 38 fifures,14 tables and 60 references.