Research And Application Of Facial Expression Recognition Based On Convolutional Neural Network

Expression recognition is a cross-disciplinary field spanning artificial intelligence,neurology,and computer science.It has a wide range of applications in psychoanalysis,clinical medicine,and vehicle monitoring.However,due to the difficulty in extracting feature recognition features and the high complexity of expressions,traditional machine learning algorithms tend to ignore features that are important for classification,resulting in lower accuracy of expression recognition.In recent years,with the development of deep learning and parallel computing,the application of deep learning-based convolutional neural networks to expression classification has attracted the attention of scholars.This is because the feature extraction of deep learning is to update the iterative weight through back propagation and error optimization,and to extract key points and features that are easy to ignore.However,there are still problems such as complex implementation,relying on a large number of data sets,and large computational complexity.Therefore,this paper focuses on the above problems to optimize the convolutional neural network and its internal structure.The main work and innovations of this paper are as follows:1.The current convolutional neural network has insufficient extraction of subtle features of expression recognition and has long training time.In order to solve this problem,this paper optimizes the GoogLeNet network structure and proposes a convolutional neural network based on parallel structure.The core is two parallel convolution pooling structures,which can extract different features from the emoticons and then feature fusion.This structure increases the width of the network and the adaptability to the scale,and the features of the image can be extracted from different angles.2.The ReLU activation function maintains the sparsity of the function but does not alleviate the problem of mean shift and neuron death.Aiming at this problem,this paper improves the traditional ReLU activation function and proposes a new parameter modification activation unit D-ReLU with two parameters.D-ReLU uses the Log function to preserve the value of the negative half-axis,and the design uses two parameters to allow the network to learn autonomously,making the output mean close to zero.Therefore,the D-ReLU function can converge more quickly,avoid the phenomenon of neuronal death,and can autonomously adjust during the network training process.3.Applying the above model algorithm to the field of Internet education,realizes theexpression recognition system for classroom video real-time classification.Use python crawler to get classroom emoticons from Baidu,Google and other websites.After manual screening and pre-processing,a small classroom expression data set is established.The data set is sent to the above model for training,and the trained model is used to identify the expression images in the video in real time,and the two educational platforms of the student end and the teacher end are built.