Research On Facial Expression Recognition Based On Static Images

Facial expression recognition,as an important research focus of affective computing,has shown broad application values in many fields.Although there are many researches based on facial expression recognition,it still faces great challenges.In the expression recognition algorithm based on traditional machine learning,a single feature extraction method gets less expression information,which may affect the final recognition performance.In the deep learning methods,the increase of depth and width in network makes the network parameters increase and the data redundant.The effective extraction of facial expression features is a key method to improve the facial recognition effect,and it is also a problem that needs to be urgently solved in this research field.Based on the above problems,this paper conducted the following research:(1)Local binary mode(LBP)feature and weber local descriptor(WLD)feature are two typical texture feature description operators,which can well describe the slight changes of facial expressions,but their focuses are different.The LBP feature focus on considering changes in pixels around the central pixel.The WLD feature mainly considers the excitation intensity and gradient direction of the central pixel and surrounding pixels.In view of the limitations of these two features,this paper combines LBP features and WLD features into new LBWP features.After extracting the LBWP features,the SVM classifier is used to classify the expressions.Experimental results show that the LBWP feature is superior to the LBP feature and the WLD feature in facial recognition,and the experimental results in the CK+ dataset and JAFFE dataset are respectively improved by 2.46% and 5.20%.(2)In the deep learning algorithm of facial expression recognition,for the problem of large number of parameters and data redundancy,this paper uses a dual-path network model to classify expressions.The dual-path network model is a combined network based on the Res Ne Xt model and the Dense Net model.The network model can not only extract deeper expression features,but also discover and learn new features during network training.Perform network training on the CK+ extended data set to avoid over fitting the data.The training network adopts 3 DPN modules,and each DPN module contains 6 micro PDN modules.The DPN model reduces network parameters,avoids data redundancy,and shortens training time.CK+ data set and SFEW data set were used to test the DPN model,and the test results were 95.3% and 57.06%,respectively.