Nuclear Norm Based Classification For Occluded Image

As an important research content in pattern recognition,face recognition has been widely concerned by researchers in the last few decades.Linear regression-based classification has become a hot topic among face recognition methods for its intuitive idea and interpretable model.Through different constraints on the representation coefficients and representation residual,linear regression-based classification is capable of dealing with recognition prob-lems under different scenarios,so it is more flexible and robust than other methods.How-ever,the existing linear regression-based classification methods do not consider the geo-metric structure in image samples,which leads to decreases in classification performance when the image sample contains a large percentage of area occluded or the training set is contaminated.In this paper,based on the intrinsic low rank property of face image,the enhanced nuclear norm based matrix regression classification algorithms are proposed.The main contents of research are as follows:(1)For the problem that the existing methods cannot guarantee to obtain accurate recon-structed image when the image sample contains a large percentage of area occluded or the training set is contaminated,an enhanced nuclear norm based matrix regression classifica-tion is proposed in this paper.The proposed model imposes a low rank constraint on the reconstructed image by using the nuclear norm,which ensures the structural characteristics of the reconstructed image,suppresses the impact of occlusion and other noise on the re-construction and improve the accuracy of classification.Meanwhile,an alternative renewal scheme is proposed to optimize the model.Experiments are conducted under the scenarios of occlusion and illumination changes.Extensive experiments results show that the proposed model is more robust comparing with other state-of-art methods,and its advantage is more evident when the image sample contains a large percentage of area occluded or the training set is contaminated.(2)For the problem that linear regression-based classification methods will arise the issue of local aliasing,which leads to the instability of the algorithm,because of the high simi-larity in human face images,a sparse nuclear norm based matrix regression classification is proposed in this paper.The proposed model embeds the sparse representation into the low rank regression to enforce the learning of the representation coefficients.The sparse rep-resentation concentrates most of the energy of the representation coefficients in the correct class,and suppresses the impact of redundant information in the data set,which helps to im-prove the discriminative ability of the algorithm.At the same time,this paper will explore the role of sparse representation and collaborative representation in recognition task with occluded images.Experiments are conducted under the scenarios of occlusion and illumina-tion changes.Extensive experiments results show that:when the percentage of the occluded area is large,collaborative representation can help to make up for the deficient information caused by occlusion so as to get higher classification accuracy;when the percentage of the occluded area is not large or the occlusion is sparse,sparse representation can suppress the interference information and concentrates the representation coefficients in the correct class,leading to better classification performance.