Research On Structural Features And Multi-view Learning Based Image Classification

Image classification is an important research topic in machine learning,and its key technology is to extract the feature representation of images.However,the real-word image data is complex,and easily affected by illumination changes,occlusion,disguise and ect.Thus,how to extract effective image features to boost the classification performance has become one of the current research difficulties.In recent decades,many scholars have presented many supervised classification models based on matrix regression,but there are still some shortcomings in these models.For example,the existing nuclear norm based matrix regression models primarily have the following defects: 1)the corrupted test samples are directly used in the regression process,which affects the recognition performance;2)the convex nuclear norm is harnessed to charaterize the low-rank structure of the error sample matrices,which may lead to suboptimal solutions.In addition,with the advancement of information technology,multi-view data is ubiquitous in real application.Compared with single-view data,multi-view data is richer and more useful.However,these large amounts of multi-view data often have no label information or only with a small amount of label information.Manual marking these data is both expensive and impractical,and even the obtained labels may be unreliable owe to some subjective reasons.Therefore,we consider carrying out clustering or semi-supervised classification on these data.Due to the curse of dimensionality as well as the limitations of multi-view denoising and fusion technology,clustering or semi-supervised classification of high-dimensional multi-view data is still a challenge.However,the existing graph learning based multi-view projection models only learn a common graph or a shared low-dimensional embedding matrix,which fails to preserve the flexible manifold geometry of each view.In view of the above background,based on sparse and low-rank theories,this paper makes an in-depth study on supervised image classification in the presence of noise with different structures.Meanwhile,from the perspective of feature selection,dimension reduction,manifold structure learning and self-weighted graph fusion,clustering and semi-supervised classification of high-dimensional multi-view data are studied.The main works are as follows:(1)A robust matrix regression model based on low-rank structure is proposed.Although most of nuclear norm based matrix regression approaches have yielded encouraging results in addressing imagewise noise,but it may lead to unreasonable classification,especially when test images are seriously corrupted by larger occlusions or severe illumination variations.This is because that they used the corrupted test images in the regression process directly,and the influence of noise will be unavoidable.To solve this problem,the corrupted image is decomposed into the sum of latent pattern and structural noise,and the recovered clean latent pattern is exploited to replace the damaged test image for regression.Meanwhile,since the latent pattern is essentially obtained by removing lowrank structural noise from the test image,it has less noise.In view of this,nuclear norm is utilized to describe the error between the corrupted test image and its corresponding latent pattern to characterize the low rank structural information of noise,and Frobenius norm is employed to depict the difference between the latent pattern and the regression image.A series of experiments well manifest that the proposed models have obvious advantages in dealing with noise caused by occlusion and illumination changes.(2)Two sparse and low-rank structure based robust nonconvex regression models are presented to deal with structural noise and mixed noise respectively.Existing nuclear norm based matrix regression models exploited the nuclear norm to characterize low-rank structural information of residual images,which may lead to suboptimal solutions and poor performance.To circumvent this problem,a parameter-free nonconvex function is used to describe the low-rank structure.The relaxation function can penalize smaller singular values to a greater degree while penalizing larger singular values to a lesser degree,accordingly,it can efficiently approximate the rank function,and the latent pattern is introduced into the regression process to improve the recognition performance.Additionally,a simple and effective classifier is designed using the nonconvex relaxation function,and an effective iterative algorithm is devised by integrating the iteratively reweighted least squares method and the alternating direction method of multipliers to optimize our models.Numerous experiments demonstrate the superiority of the proposed models in handling structural noise and mixed noise.(3)A multi-view learning framework based on projection and consensus graph is developed for clustering and semi-supervised classification,which performs feature selection,manifold structure preservation,consensus structured graph learning as well as data clustering/label propagation simultaneously.Graph-based learning approaches have achieved remarkable success in clustering and semi-supervised classification of multi-view data owing to the capacity to reveal the relation between data and discover its underlying structure.However,the real multi-view data is beyond simply high-dimensional,but always contains noise and redundant information,so that the learned affinity graph from this type of data may be unreliable,resulting in inaccurate results.To alleviate the problem,2,1-norm is leveraged to constrain the projection matrix to resist noise and adaptively select discriminative features.Moreover,to preserve the local manifold structure information of all views and make the best of the consistent and complementary information among multiple views,we construct an informative similarity graph for the projection data ofeach view,and adopt a auto-weighted strategy to fuse them according to the importance of different views.In addition,for clustering task,the Laplacian matrix rank constraint is introduced to make the learned consensus affine graph have an explicit clustering structure,which directly presents the cluster label of each data point without any subsequent clustering steps.Numerical studies on several multi-view benchmark datasets justify the superiority of the proposed approach over the state of the arts.