Research On Facial Expression Recognition Based On Deep Learning

With the rapid development of pattern recognition and graphic image,facial expression recognition(FER)has gradually become a key technology in intelligent and diversified human computer interaction.The existing popular FER models mainly apply facial expression recognition to deep learning,to avoid such problems as tedious training process and single artificial design features in traditional expression recognition.However,this method still has some limitations,such as: when face expression recognition is in a complex environment,face images have different face poses and different light intensity;When using the integrated model to improve the recognition rate,the training parameters and time of the model increase exponentially.When in the practical application environment,it is impossible to collect a large number of label samples in time to participate in the training.To solve this problem,this paper deeply studies the facial expression recognition methods based on deep learning,and the main research work is as follows:1.In order to improve the recognition rate of FER model,this paper firstly proposes a face expression recognition method based on LBP and CNN,which mainly includes two branch networks: The other branch is the deep global feature network based on CNN,which mainly extracts the global feature of image;Another branch is the depth texture feature network with Local Binary Pattern feature graph as input,which mainly extracts the local texture feature information of the image.Finally,on the CK+ JAFFE classic dataset and multi-database,the recognition rate of the model can reach 98.23%,95.24% and 93.99%,respectively.Three different experiments were conducted to verify the effectiveness of the method.2.Aiming at facial expressions in complex situations,a FER method based on depthwise separable convolution is proposed.Firstly,the image regions that are most relevant to facial expression recognition are segmented by image preprocessing to reduce the impact ofcomplex environment;Secondly,two basic classification models are constructed based on the VGG19 and Resnet18 network structures,and the conventional methods in the model are replaced with deep separable convolution Convolution to reduce the weight of the model;Finally,two pre-trained base classification models were fused by joint fine-tuning.Finally,the experimental results show that the recognition rate on the FER-2013 data set reaches 75.15%,which is 3.95% higher than the Kaggle Expression Recognition Challenge champion;the recognition rate of CK+ and JAFFE data set were as high as 98.98%、97.14%,At the same time,based on the model,the superiority of the combined fine-tuning and weighted summing fusion method is compared.3.In order to reduce the workload of manual labeling,based on the idea of improved co-training algorithm,a facial expression recognition method based on semi-supervised learning is proposed.Firstly,use two different deep learning networks to build and pre-train two base classifiers under the same labeled attribute set.Then,use different base classifiers to classify and predict unlabeled samples,and then use JS Degree to calculate the difference between different distributions.the model is trained using a small number of labeled samples and a large number of unlabeled samples simultaneously.Finally,on CK+ and FER2013 data sets,the training sets were divided into proportions of 0.2,0.3 and 0.4 to verify the effectiveness of the method through recognition rate.