Study On Face Recognition Based On Collaborative Representation

With the development of computer technology and biometrics,biological features such as face,fingerprint and iris have been widely used in the field of identification because biological features are unique,easy to carry,not easy to lose and so on.Among them,the face has the advantages of non-contact,easy acquisition and so on,and it has been maintained a high degree of research in the field of identification.Recently,the collaborative representation based classification has been applied to the field of face recognition,although it makes up for the low accuracy of the sparse representation based classification for small sample,it is less robust to face images with occlusion or illumination changes.Second,there are many similar images in different face images in face recognition,so the correlation between dictionary vectors is very high,which leads to the final classification error of collaborative representation based classification.Therefore,this paper improves on collaborative representation based on classification to improve its classification performance.At the same time the feature used is the original gray value,the feature which contains weak image information is not conducive to the final classification.Therefore,this paper introduces a feature extraction method to extract the features that can reflect the face information.This paper studies two aspects of facial feature extraction and classifier.The main research contents are as follows:1)Illumination changes and occlusion are inherent problems in face recognition,the paper proposes class specific dictionary learning for local kernel collaborative representation based classification(CSDL-LKCRC).Firstly,the face images are divided into block images,each sub-block is mapped to a higher dimension space by Gaussian kernel.Then,each sub-block is combined with class specific dictionary learning to get the corresponding reconstruction error based on local kernel collaborative representation.Finally,according to the reciprocal of the reconstruction error,the process from local discrimination to the global classification is completed by the voting.Experimental results on Extend Yale B?AR?CMU PIE face database and mixed face database(AR,Extend Yale B and CMU PIE face database)show that the recognition accuracy rate is 99.8%,98.8%,93.9%,87.1 %,respectively,compared with the recent CSDL-CRC method,increased by 10.4%,7.5%,4.6%,8.2%.The experimental results show that the proposed method not only has high recognition accuracy,but also has strong robustness to illumination changes and occluded face images.2)Asymmetric Local Binary Pattern(AR-LBP)can't fully express face information and collaborative representation based classification(CRC)algorithm introduces inter-class interference,therefore,this paper proposes the combination of multi-layer AR-LBP features and Weber Local Descriptor(WLD)features to complement the features,and the interference between classes is reduced by increasing the sparsity in CRC.Firstly,the multi-layer AR-LBP features of face images are extracted and the AR-LBP features of all layers are cascaded.Secondly,multi-layer AR-LBP-WLD features are obtained by cascading multi-layer AR-LBP features and WLD features of original face images.Finally,face classification is done using sparsity augmented collaborative representation based classification(SA-CRC).In order to verify the specific performance of the proposed method,this paper carried out simulation experiments on ORL face database,Yale face database and GT face database.On the three face databases respectively,the three types of feature extraction and WLD feature,combined with CRC classifier and SA-CRC classifier proposed in this paper,are used to complete the face recognition experiment.The experimental data showed that the recognition accuracy of AR-LBP-WLD increased by 0.9%~42.6% compared with that of AR-LBP,WLD,LBP-HOG and multi-layer LBP-HOG.When SA-CRC is used instead of CRC,recognition accuracy increases again by 0%~0.8%.The experimental results show that multi-layer fusion method more fully characterizes the face features and Multi-layer AR-LBP and WLD feature fusion better characterize the face features,and the replacement of CRC by SA-CRC can further improve the recognition accuracy.In summary,this paper improves the overall face features in face database and single face features respectively based on the collaborative of class specific dictionary and collaborative of fusion multi-layer AR-LBP and WLD.Comprehensive experimental data show that the collaborative representation is obviously conducive to the improvement of face recognition performance;At the same time,the research results of this paper also show that the collaborative of facial global features and individual local features should be a beneficial research direction.