Research On Robust Low-rank Representation For Images Based On The Evaluation Of Features And Samples

With the popularization and application of the Internet and big data,a large number of image data with different shooting environments and different quality have been produced.Effectively realizing the classification and recognition of these images has important theoretical significance and practical value.The representation of image features plays a vital role in image classification and recognition.Besides,the representation of images has always been an important problem and mission in the field of computer vision and pattern recognition.The main problems faced by the representation learning include the interference caused by noise,occlusion or loss in the image,the dimension curse caused by the dimension increasing,and the difference and identities of multi view images.To address the mentioned problems above,this dissertation investigates the correlations among image features and the low-rank structure to learn the potential relationships among features and typical projection vectors,the latent relevance of features and residual projections,and the connections between samples and features for low-rank data fitting respectively.A series of robust image low-rank representation methods based on feature and sample evaluation are proposed.The main contributions of this dissertation are summarized as the following 3aspects:(1)A robust image representation method based on feature evaluation and canonical correlation analysis is proposed.Aiming at the problems of noise interference and high feature dimension in image feature learning,a robust image representation method based on feature evaluation and canonical correlation analysis is proposed.In this method,the feature factor matrix is introduced to measure the proximity between each feature vector and the main projection vector in feature dimensions,and evaluate the contribution of each feature to the overall feature space,so as to realize the evaluation of features.Therefore,different features can obtain corresponding weights to suppress noisy data.In order to achieve a more comprehensive evaluation of features,it is further extended on this basis.By introducing the matrix deflation mechanism,multiple factoring matrices for multiple projection vectors are constructed,and the importance of each feature in each projection is weighed respectively,so as to obtain more robust image feature representation.The experimental results on several public data sets show that the proposed method has low computational overhead and can effectively improve the identifiability of features and significantly improve the accuracy of image classification.For example,it improves 1.57% ～3.78% and 1.69%～ 2.47% on ORL and COIL-20 respectively,and it also has better anti-noise performance.(2)A robust low-rank image representation method based on feature evaluation and residual projection is proposed.Aiming at the problems of noise and occlusion in image data,a robust image low-rank representation method based on feature evaluation and residual projection is proposed.Different from the traditional low-rank representation method,which regards the residuals as noise data or outliers,the proposed method tries to learn robust features by modeling the structure of residuals.That is,the residual is regarded as a measure of the distance between the input data and its low rank representation.Therefore,the proposed method first learns the robust low rank projection from the residual,then finds the good matching structure between the input data and its low-rank representation to enhance the weight of highquality features in the feature representation.So as to reduce the dimension and suppress the interference of data noise and image occlusion at the same time.The experimental results on public image data sets such as ORL,YALE,Yale B,AR,LFW,COIL20,COIL00 and Caltech101 show that the proposed method can obtain higher classification accuracy than the current state-of-the-art(SOTA)methods in complex image classification tasks.For example,the classification accuracy of LFW and caltech101 is improved by 0.42% ～7.91% and 1.42% ～5.36% respectively.The feature representation has better robustness on noisy and obscured image datasets as well.(3)A robust low-rank representation for multi-view image based on feature and sample evaluation with residual projection is proposed.Aiming at the consistency of multiview image feature representation,a robust multi view image residual projection low rank representation method based on feature and sample evaluation is proposed.In this method,the residual projection model in single view is extended to multi-view scenarios,and an image lowrank representation model of multi-view residual projection is proposed.The model optimizes the overall model by sharing the same low rank structure among multiple views to learn the consistency characteristics form different views,so as to improve the identifiability of multiview image representation.In the model of low-rank representation of multi-view residual projection,the sample evaluation mechanism is further introduced to construct the sample evaluation matrix.In this model,the contributions of different samples can be learnt via residual projection model.It can strengthen the weight of the samples with higher discriminatability and weaken the weight the samples with lower discriminatability,so as to realize the robust low rank representation of multi view images.Eventually the proposed method is applied to multiview image classification.The comparative experiments on multiple public databases show that the proposed method can effectively improve the accuracy of multi view image classification compared with other advanced feature representation methods.For example,on multi-PIE and caltech101,it is 0.95%～ 2.64% and 2.11%～ 4.29% higher than the comparison methods respectively.The anti-noise performance is also improved significantly.