A Study Of Sparse Subspace Clustering For Image Sequence And Its Applications

With the advent of the new generation of artificial intelligence,the current research in the field of computer vision is faced with several basic problems,such as the extremely high dimension data,the large number of data samples,the nonlinear data lacking the prior knowledge of the internal low-dimensional structure,and at the same time the video data's easily pollution by noise.How to analyze these types of data effectively and construct an efficient clustering algorithm is one of the difficulties to be overcome in the field of machine learning,which is full of importance and practical application value.Although the sparse subspace clustering method provides a good solution for these types of data analysis,there are still some problems at present.For the nonlinear image sequence data,current sparse subspace clustering algorithms simply regards the image as pixels pooling,which ignores the description of the essential spatiotemporal attributes of the image data,so it is difficult to fully express its semantic information.This paper studies the sparse subspace clustering methods based on spatiotemporal characteristics to excavate the spatiotemporal law of image sequence.Based on the general representation model of sparse subspace clustering algorithm,this paper comprehensively considers the order of image data in time domain and space domain,and it then proposes a set of improved sparse subspace clustering algorithms.The contribution of this paper is as follows:It is difficult to characterize the spatio-temporal essential properties of nonlinear ordered data in full mode,at the same time this SMR model adopts the K nearest neighbour(KNN)graph to select samples for representation.All neighbours in the KNN graph are assumed to be equally important candidates.Firstly we use special weights that are calculated by the kernel random walk and a novel cross-view kernel function to evaluate the contributions of neighbours to the subspace clustering in SMR.The neighbours that are found by the Gaussian similarity formula can be considered long-range similar neighbours.We add another item to accurately reflect the order relation in the cross-view kernel function.This addition allows the kernel function to generalize the conventional SMR method for sequential data.Furthermore,we also proposed a novel method for automated segmentation on the mobile video scenery.It uses the kernelized random walks on the globe KNN graph and the Smooth Representation Clustering to improve the segmentation effectiveness.The high order transition probability matrix of the kernelized random walks is utilized for erasing the unreliable edge of the graph.Aiming at the poor effect of sparse subspace clustering in high-noise environment,the article uses wavelet multi-scale transform to mine the spatial prior information of image data.Wavelet transform reveals the internal structure of the image,reflects the continuity of the pixel space of the image,and achieves the effect of noise suppression.In view of the damage to the block diagonalization structure of similarity matrix caused by the random selection of sparse coding coefficients in sparse subspace clustering method,this paper use de-noising Laplace constraint to promote the maximization of intra-class correlation and the minimization of inter-class correlation.At the same time,inspired by Tierney's subspace clustering of the sequence data,a novel subspace clustering method based on sequential character is proposed.In the beginning the lifting wavelet transform is applied to exact low-frequency information of the signal,and then a stronger special penalty is applied to emphasize the similarity between adjacent samples,in which the penalty factor is automatically adjusted according to the noise without manual intervention.Considering the obvious graph features and probabilistic structure properties of highdimensional image data,the article uses the wavelet-HOG transform in the kernel view to extract both the graphical features of image(with the wavelet process)and the probability and statistics properties structure(with the HOG process)from the image.In addition,we assign different weights to different features to obtain a sparse coefficient matrix that helps to emphasize the global and local correlations in each sample.We also describe the timeaxis attributes of the video on this basis,and propose the sequential subspace clustering method(Lp OSC)based on the penalty term of Lp norm to reflect the temporal correlation of adjacent frames in the video.This is a generalization and improvement of the image sequence subspace clustering model.In summary,this paper proposes sparse subspace clustering methods based on spatiotemporal characteristics,which can be applied to the clustering analysis for images and video data without labels.After a set of experiments,the above algorithms show that they can not only enhance the effectiveness,but also partly solve the image data of high noise,lighting problems,deformation problems and angle problems to improve the accuracy and robustness.In the future,we will study how to improve the speed of these methods with distributed optimization technology.