Analysis And Research On Key Techniques Of Facial Expression Image Recognition

Expression recognition technology is a cross product of many subjects such as biology,psychology and computer science.Because of its high use value and research significance in realizing intelligent human life,it has been widely concerned by many researchers in recent years.As it is an advanced technology covering a wide range and involving many fields,it is accompanied by many technical problems that need to be overcome.In this thesis,we study in depth on the key issues such as extraction methods of facial expression features,feature classification and recognition,occluded facial expression classification and recognition.We also analyze the deficiency of current methods and propose related solutions which are proved effective and feasible by theoretical analysis and experiments conducted.The research emphases and innovative points of this dissertation are as follows:1.Facial expression recognition based on strong separability feature selecting function and discrete Shearlet transform.In view of the problems of basis function support interval and finite decomposition direction in feature extraction of facial images by wavelet transform,a discrete Shearlet transform is proposed as a method for facial feature extraction.By analyzing the advantages and disadvantages of Shearlet transform decomposition coefficients as expression features,and the performance of Shearlet transform coefficients in the low-frequency and high-frequency domain,a function is proposed to evaluate the strong separability of low-frequency and high-frequency decomposition coefficients.This function can filter the characteristic coefficients.The optimal Shearlet coefficients can be selected by measuring the classification ability of the coefficients at different decomposition scales and directions.The final feature of expression classification is obtained by fusing the low-frequency and high-frequency coefficients under the optimal decomposition scale and direction.The optimal parameters of Support Vector Machine are selected through experiments,and the output of expression classification results is completed by designing multi-classifier framework.The feasibility and effectiveness of the proposed method are proved by simulation experiments and comparative experiments.Compared with existing methods of facial expression recognition,this method has the following advantages.Firstly,as a coefficient extraction method,the discrete Shearlet transform can be more effective in capturing the changes of facial features in facial images because of its multi-scale,multi-direction,simple mathematical structure and fast implementation.Secondly,this method only selects the coefficients with the best classification ability as the expression features,which can ensure the recognition accuracy and reduce the data amount of the feature matrix,so as to make the classification and recognition of expression more efficient.Finally,the low-frequency coefficients reflecting the general image and the high-frequency coefficients reflecting the changes of the expression details.The two are fused as the expression features in this method,which provides a guarantee for high precision identification.2.Facial expression recognition based on discrete Shearlet transform and normalized mutual information feature selection.Aiming at the problem that the expression image coefficients after discrete Shearlet transformation have large amount of data and redundancy,a method for selecting initial feature coefficients is proposed.By analyzing the feasibility and deficiency of feature selection method in feature coefficient selection,an improved normalized mutual information feature selection method is proposed to find the optimal feature subset of the original feature coefficients set.By optimizing the difference between the original features and categories mutual information and features mutual information,the selection of optimal feature subset can be completed,and the key classification information of original feature coefficients matrix can not be lost.Furthermore,t-distributed stochastic neighbor embedding nonlinear dimension reduction is carried out for feature subsets to ensure the simplification of final classification features.Finally,multi-classification Support Vector Machine is designed to classify and recognize facial expressions.The visualization results of classification features,recognition rate and comparison experiments all prove the feasibility and effectiveness of the proposed method.Compared with the existing methods,our method can simplify the feature coefficient matrix with large data and redundant information,and reduce the feature coefficient dimension and ensure the high precision recognition,saving the running time and simplifying algorithm complexity.Therefore,our method can provide a strong guarantee for efficient and accurate recognition.3.WGAN-based occluded facial expression recognition.In order to solve the problem of occlusion in expression images in real life,a method based on Wasserstein Generative Adversarial Network is proposed to complement the occlusion region in occluded expression images.By analyzing the shortcomings of Jensen-Shannon divergence as similarity measure standard in Generative Adversarial Network,Wasserstein distance is proposed to measure the distance between two probability distributions.A generator G and two discriminators_iD,i?28?1,2are used to construct the network structure of this method,and the generator fills the occlusion expression images naturally under the triple constraints of weighted reconstruction loss,triplet loss and adversarial loss proposed by us.Furthermore,Wasserstein distance is used to construct the adversarial loss between generated images,original un-occluded images and occluded images in irrelevant regions to optimize the real and fake discrimination ability of discriminatorD₁.At last,facial expression classification recognition is completed by introducing classification loss to discriminatorD₂.The effectiveness of our method is proved by visual occlusion filling results and the comparison of expression recognition rate with/without filling.Compared with the recently methods,our method is an improved Generative Adversarial Network and it can complement the occluded area in the occluded images,so the complementing images can be obtained.The generated de-occluded image has complete face information,thus ensuring the smooth progress of subsequent expression feature extraction and classification.And after effectively filling the face occlusion image with complex expression information,the reduced and most divisible set of expression feature coefficients can be obtained,and the processed image quality is higher and the recognition is more accurate.