Research And Verification Of Facial Expression Recognition Based On Deep Learning

Facial expression recognition is an important branch in the field of computer vision(facial recognition).Many factors including the diversity of fical expressions,changes in head posture and the environment,bring a big challenge to the work of facial expression recognition.In this thesis,focused on the related researches in the field of FER,we have done the following works:(1)The traditional convolutional neural networks need the large number of model parameters and the accuracy of facial expression recognition can't be improved explicitly compared with traditional machine learning algorithms,aimed at this condition,firstly,we strat with the lightweight models,studying the lightweight structure of the model and the optimization process of network training.Concretely speaking,we use a combined training strategy approach to train the three excellent models(VGGNet,Mobile Net,and Res Net)with pre-training,fine-tuning training,and so we can give the network model a better initial value and make the weight parameters reach the ideal value faster,and obtain better accuracy,time and computing resources saving.Secondly,combined with the network structure of these excellent models,an improved convolutional neural network model based on a deep separable convolution structure is designed,we use the model to conduct the experiment on dataset FER2013,compared with the traditional convolutional neural network,the results show that the network structure of the model is optimized,the number of model parameters is greatly reduced,and the utilization rate of model parameters is improved under the condition of ensuring a high accuracy rate of 68.31%.Finally,we further study the impact of model light weighting-model pruning on the model,applying Mobile Net as the benchmark model,after weitht pruning,the model's parameter compression rate reaches to 65.94%,the accuracy of the model does not change significantly,but the model parameters utilization efficiency has been further improved.(2)The existing reliable data sets are small in scale and the(data quality)reliability of large-scale data sets is low,in light of that,we start with multi-task learning,studying the impact of multi-task learning and its loss function on facial expression recognition tasks.First,we still use VGGNet,Mobile Net,and Res Net as experimental comparison models and use the FER2013 data set for pre-training,and the CK+ dataset for fine-tuning,and achieved a higher accuracy.Second,giving a pair of training images,we predict the categories of the two images and judge whether they belong to the same category.Finally,selecting Res Net,which has the best recognition effect on the CK+ dataset,as the intermediate sharing model for multi-task learning,and we do experiments on the weight of two loss functions(linear addition loss and weighted sum loss).The experimental results show that when we use the second loss function,on the CK+ dataset,we get an accuracy of96.5%,on the FER2103 dataset,we get an accuracy of 72.26% and on the SFEW2.0 dataset we get an accuracy of 53.61%.