Research On Expression Recognition Algorithm Based On Deep Learning

Expression recognition technology is a hot and difficult problem in the field of machine learning,and has a good application prospect in many fields.Fast and accurate recognition of human facial expressions can meet the growing social demand.However,there are still different problems with the current recognition technology: in the feature extraction stage of the expression recognition model,due to the short time of expression occurrence,the small magnitude of the action,and the different magnitude of the expression action occurring in different facial regions,if more detailed expression features are not extracted,the recognition rate of the expression model will be low and the generalization ability is insufficient.In the training phase of the expression recognition model,due to the inadequate sample size,unbalanced sample categories and the difference in distribution between the training and test sets,the expression recognition model is weak in recognizing individual classes of expressions and has poor robustness.In order to solve the above problems,the work and results of this thesis are as follows.(1)A multi-scale feature extraction depth model(FCNet)for expression recognition is designed.FCNet consists of two parts-a multi-scale expression feature extraction module(Fef)and a coordinate attention mechanism algorithm.The Fef module mainly consists of three Tfe structures.Tfe is a three-branch feature extraction structure,which uses convolutional kernels of different scales,a BN layer after each convolutional layer,and a non-linear variation of the Relu activation function to enhance the generalisation ability and expression feature learning ability of FCNet and improve the training speed and convergence speed.After the feature extraction was completed by the Tfe structure,the FCNet was made to acquire features incorporating both shallow and deep information by means of a jump connection.This was followed by the use of a coordinate attention mechanism to assign channel weights based on face location information to these fused features,enhancing FCNet’s ability to learn from expression data and ultimately improving the performance of expression recognition.(2)In order to continue to improve the performance of expression recognition,an oversampling algorithm,the FM algorithm,is firstly designed.Secondly,based on the FM algorithm,the FM-Meta pseudo-labelling algorithm is further designed.the FM-Meta pseudo-labelling algorithm uses two types of data to train the expression recognition model-unlabelled data and labelled data.the FM-Meta pseudo-labelling algorithm contains two neural networks--the teacher network and the student network.In contrast to the pseudolabelling algorithm,the teacher network adaptively adjusts to feedback from students’ performance on the labelled data set,based on the difference in the distribution of labelled and unlabelled data.The updated teacher network guides the student network in the form of pseudo-labels generated based on unlabelled data.By training the teacher network in parallel with the student network,the expression recognition performance of the student network is improved and made more robust.Experimental validation was carried out on the FER2013 and MMA datasets.The experimental results indicate that FCNet shows better robustness in expression recognition accuracy compared to traditional algorithms.The further use of the FM-Meta pseudolabelling algorithm not only improves FCNet’s expression recognition,but also provides a new training method for the model to handle unlabelled data.