Research On Facial Expression Recognition Based On Improved Convolutional Neural Network

With the development of machine vision and pattern recognition,facial expression recognition has gradually become an active research topic.It has a wide range of applications in clinical medicine,psychological analysis and automobile driving monitoring.In practical applications,facial expression recognition still faces many challenges due to uneven light,low resolution,noise and occlusion.Facial feature extraction is particularly important as a key technology of facial expression recognition.Traditional feature extraction methods tend to ignore some important features,which leads to unsatisfactory recognition accuracy.With the rise of deep learning,convolutional neural networks have attracted the attention of many scholars.Through backpropagation and error optimization iterative weights can extract the characteristics of key points.However,the complexity of calculation and dependence on a large number of data sets also bring us inconvenience.This article mainly optimizes the convolutional neural network and the analysis of the internal structure for the above problems.The main work and research content of this paper:(1)We propose an improved convolution model for neural networks in order to solve the problem that the recognition of facial expressions is too complex to be calculated with conventional convolutional neural networks.Initially,a nested group convolution structure was used.This effectively reduces the number of connections between network layers.At the same time,the interleaved structure also allows a certain amount of information communication between different coil units.Since the front end of the network has a strong negative correlation,the cascaded linear rectifier unit(CRe LU)is used as the activation function in the initial stage.The intermediate optimization stage combines the deep separable convolution with the residual network to further optimize the network feature extraction performance.Finally,add the Softmax layer to obtain the classification results.Detailed experiments were carried out on the Jaffe,CK + and Fer2013 data sets,and the accuracy rates reached 95.45%,98.25% and 72.23% respectively,and compared with the proposed algorithm,it fully verified the feasibility of the method.(2)Aiming at the complicated problem of extracting facial expression features under unrestricted conditions,an algorithm that combines global abstract features in deep learning with local texture features extracted by conventional methods is proposed.Using circular LBP as the basic operator,the feature statistics in the image are improved through the equivalent mode,and the problem of too many binary modes in the image is solved.The stable feature of the image is extracted by the rotation invariant operation,and finally the LBP is extracted The features and the features extracted by the deep network are processed in parallel to obtain the final result.The accuracy rate of the Fer2013 data set reached 73.31%,an increase of one percentage point.(3)Apply the above algorithm to patient status monitoring,mix the Jaffe data set and CK+ data set in the experimental environment to obtain a new mixed data set,send the preprocessed data set to the convolutional neural network for training,and use the training The model performs real-time detection and recognition of sequence frames in the video,and real-time analysis of expression classification results,the facial expression recognition system for real-time status monitoring of remote classroom students is realized.