Subspace Learning Via Random Sample Probing And Robust PCA And Its Application In Video Processing

Video processing is one of the most basic tasks in artificial intelligence,and its results directly affect subsequent image recognition,image classification,behavior detection and so on.A large number of studies show that the high-dimensional video data is not disordered.Through the study we find that the high-dimensional data exist in the low-dimensional subspace.Subspace learning is a widely used and effective technique in video analysis and has been successfully applied in video compression,target recognition and so on.We mainly study the four problems of subspace segmentation,subspace recovery and subspace compression in video processing.Our contributions are as follow:The current subspace segmentation method based on spectral clustering,such as:Sparse Subspace Clustering(SSC)can not satisfy the segmentation of large data points,the time complexity is O(n³).For fast segmenting large scale data,this paper proposes a fast subspace segmentation method called Random Sample Probing(RANSP),which randomly selects seed points and uses Ridge Regression(RR)to calculate the correlation between other points and this point,so that you can get a subspace.The Wood Bury formula is used to optimize the RR solution so that the time complexity of RANSP is linear,so large-scale data points can be processed quickly.Robust principal component analysis(RPCA)aims to recover low rank and sparse parts from polluted data,which has been widely used in computer vision,image processing and so on.However,there are still areas could be improved in RPCA model.In order to improve the subspace recovery ability of RPCA model,we propose a Truncated Weighted RPCA(TWRPCA)model.In TWRPCA model,we keep the first 10%of the singular value unchanged,and give different weights to the other singular values.At last the Inexact Augument Lagrangian Multiplier(IALM)is used to solve the TWRPCA model.RPCA subspace learning method and its extended model are successfully applied to moving object detecion,but can not deal with the slow moving target,which is easy to cause"cavity",especially in complex backgrounds,such as swaying leaves,fluctuating lakes,etc.In order to detect slow moving targets in complex environments,video segmentation constraints are used for dynamic backgrounds,and salience constraints are imposed for slow motions.Finally,this paper integrates video segmentation and salient constraints into the RPCA model.The SSC-RPCA model is obtained.Unlike other literatures,the significant information of this paper is not calculated in advance but obtained by solving the SSC-RPCA model.Auto-encoder has been widely used in image and video compression,but traditional auto-encoder needs to store a large number of network parameters.Traditional auto-encoder requires specific data to train the network,so their generalization capabilities are weak.This paper proposes a 3D Tensor Auto Encoder(3DTAE)subspace compression method,in which video is compressed into a nonlinear subspace.In the traditional auto-encoder methods,the video is represented as network parameters and vectors.In this method,the video is directly represented as a network parameter.Assuming the dimension of the input data is n,and then the dimension of the network parameter is also O(n^1/3),which can meet the needs of video compression.In addition,because the video is compressed directly into network parameters,there is no test process in 3DTAE,so there is no generalization problem.