Study Of Facial Expression Recognition Based On Partial Occlusion

The research of facial expression recognition in the controlled environment has achieved good results,but in the real environment,the facial expression recognition is often affected by occlusion,illumination,posture and other factors.It makes the real sense of facial expression recognition bottlenecks encountered.This paper focuses on solving the occlusion problem in case of facial expression recognition,and carries out the research from the following two aspects:(1)Two novel feature extraction algorithms are proposed to solve the shortcoming that the traditional feature extraction algorithms have in feature description.(2)In the aspect of image reconstruction,the traditional image reconstruction algorithm is studied and analyzed,and a new algorithm of image reconstruction based on information entropy Principal Component Analysis is proposed.The specific research works are as follows:(1)In this paper,the method of "three court five eyes" is used to detect and locate the human face,and the obtained face image is processed by Gauss filter.Then in the aspect of feature extraction,aiming to the inadequacy of original Center-Symmetric Local Binary Pattern(CS-LBP)and Difference Local Directional Pattern(DLDP)in describing the feature of the occluded image,Difference Center-Symmetric Local Binary Pattern(DCS-LBP)and Gradient Center-Symmetric Local Directional Pattern(GCS-LDP)are proposed.In order to further reduce the impact of occlusion,the face image is segmented to obtain the eye and mouth regions.And the chi square distance of image blocks are weighted by using the method of information entropy.The final recognition results are got by using the nearest neighbor classifier.The experimental results show that this method can effectively reduce the impact of occlusion.(2)With the increase of occlusion area,the error that caused by utilizing the feature of non-occluded region of the image individually becomes bigger.So considering from the perspective of image reconstruction,this paper improves the PCA algorithm and presents a new image reconstruction algorithm based on information entropy PCA,to fully utilize the feature of the occluded region.As for the noise introduced in the process of image reconstruction,this paper uses Scale Invariant Local Ternary Pattern to encode the reconstructed image.The image which is encoded by Scale Invariant Local Ternary Pattern(SILTP)is divided into the eye and mouth regions,and Local Directional Texture Pattern(LDTP)features of these two regions are extracted,in order to effectively highlight the main information in the image,and reduce the interference of the unimportant information.In terms of the classification,aiming at the eye and mouth regions of the reconstructed image,two support vector machines are used to classify.These two results and two classification results of the two regions in step(1)are fused to obtain the final recognition result.The experimental results show that the method can obtain better recognition result when the occlusion area is larger.