Image Set Representation And Classification Based On Low Rank Representation And Sparse Sample Selection Model

With the rapid development of computer science,a large amount of data information,such as images or videos,has been collected to generate massive data.How to effectively analyze and process these data is an urgent issue that needs to be resolved.Recently,the problems of image set representation and recognition have attracted many attention.Compared with single images,the processing of image set problems is more complicated.The images are collected with different images of the same sample,so the information provided is more effective and abundant.Therefore,the related issues of the image collection are the current research hotspots,and they themselves have very important application values.The low rank representation method has been successfully applied in computer vision and pattern recognition.The proper image set representation method can discover the potential internal structure of the image set,which is conducive to further analysis and processing of the image,and can remove the unwanted interference information in the image.Can extract more quality information,make the image set more accurate.For image set representation and classification problems,on the one hand,how to construct a set representation model to represent all the information in the image set while suppressing the effects of redundant information,noise data,and background information.On the other hand,similarity measurement and classification methods between image sets also need to be studied.Generally,this thesis mainly focuses on the following several aspects:(1)In real applications,the image set samples are often corrupted by various kinds of noises,corrupted data,and error information,which makes the image set recognition and learning tasks more challenging.In order to deal with this issue we propose a method to represent the image set which combines the low-rank representation with graph representation.Then,in order to obtain more accurate reconstruction,we add the optimal mean constraint term.Experiments results on several benchmark datasets demonstrate the effectiveness of the proposed image set representation and classification method.(2)Recent studies have revealed that images are usually distributed in a manifold space.Then sets can be represented by manifolds.The distance between sets can be calculated by the distance between manifolds.Inspired by the existing research,the image set representation in this paper can be incorporated into the manifold structure information.We therefore propose a new manifold regularization low-rank representation method for image set representation.The Laplacian manifold constraint is added to the objective function of the algorithm as a penalty term,so that the algorithm can maintain the manifold structure of the image samples.Experimental results prove the robustness and effectiveness of the method.(3)The samples we collected were obtained through video capture.Due to the temporal interference,the captured images contained a large number of duplicate samples.At the same time,due to illumination,occlusion and other factors,it also caused the appearance of erroneous data.The classification result will be affected.So this paper proposes a manifold regularized sparse sample selection method to obtain a representative sample of the image set.This not only sparsely reconstructs the image set,but also preserves the manifold structure information within the image set during the reconstruction process.At the same time,the selection of representative subsets can remove redundancy,error information,and noise data,making the representation of the image more accurate.After the regularized sparse sample selection process,we then use covariate relation graphs to represent representative samples in image set learning task.Experimental results on four benchmark datasets demonstrate the effectiveness and benefits of the proposed method.