Illumination And Pose Variable Face Recognition

With the progress and development of the Internet,authentication technology has been widely used in security precaution.Compared with traditional authentication technology,biometric identification technology is more convenient,stable and not easy to be forged,so it has been widely used.As one of the most important biometric identification techniques,face recognition technology has gained wide attention in pattern recognition and machine learning.Recently,constrained face recognition has good effect,but uncontrolled face recognition has huge challenges,which mainly includes variable illumination,pose,facial expression,occlusion,etc.As for variable illumination and pose in face recognition,this paper researches insensitive feature extraction for illumination and pose,the fusion of identified classifiers,frontal face synthesis.The main work is as follows:(1)Due LBP and HOG features are not sensitive to illumination and pose,and fusion features have high dimension,this paper puts forward the method via the fusion of ULBP and improved multi-scale HOG features(MHOG).Because faces in different regions contain different amounts of information,each face image block is extracted ULBP_MHOG features,and the weight of each block is obtained by using the information entropy at the same time,then adaptively weighted ULBP_MHOG features(AWULBP_MHOG)are obtained.Finally,the weighted sparse representation classifier(WSRC)is used for illumination and pose variable face recognition.Experiments prove that proposed method effectively increases illumination and pose variable face recognition rates.(2)Deep learning plays an important role in face recognition,this paper puts forward integrating deep learning features and AWULBP_MHOG features to make features complementary.Trained VGGFace model is utilized to extract VGG features of 4096 dimensions,and PCA is used to reduce feature dimension,then VGG_PCA features are obtained.AWULBP_MHOG_VGG_PCA features(AMVP)are obtained by fusing AWULBP_MHOG and VGG_PCA features.Finally,the weighted sparse representation classifier(WSRC)is used for illumination and pose variable face recognition.Experiments prove that proposed method further improves illumination and pose variable face recognition rates.(3)In order to further eliminate the influence of pose,this paper proposes a new method of WSR frontal face synthesis via extended neighborhood search and random blocks.To make up for the defect of manual face alignment,first each face image block is extened by using neighborhood search,then random blocks and discrete cosine transform(DCT)are used,which reduces elapsed time.Finally,test face is represented by weighted sparse representation(WSR)to obtain weighted coefficient of sparse representation for frontal face synthesis.The experimental results show that the quality of synthetic face image is improved,which is beneficial to pose variable face recognition.Experiments on multiple public face databases prove that the combination of AMVP features and WSRC is a good solution for illumination and pose variable face recognition.The WSR frontal face synthesis method based on extended neighborhood search and random blocks can further solve pose variable problems,which can improve pose variable face recognition rates.