Structured Sparse And Low-Rank Representation Learning Method And Application Research

Image representation learning is an important research content in the fields of pattern recognition and machine learning.It is also the most basic structural unit in computer vision and multimedia processing systems.The performance of features directly determines the performance of subsequent information processing modules.Traditional machine learning is often unsatisfactory in the face of massive,high-dimensional,and complex structured image data.The sparse model and low-rank model can effectively reveal the sparseness and low-rank of the data itself,and reveal the underlying structural characteristics of the data,thereby greatly improving the data processing efficiency and effectiveness.In recent years,it has been widely used in computer vision,machine learning and image processing.Inspired by this,this dissertation aims to mine data structures,build sparse and low-rank models,and use structured constraints to enhance the model's learning efficiency and robustness,and finally apply it to faces,biological gene and radar image feature expression and classification.The main work and innovations of the paper are as follows:1.Classical sparse representation models in classification tasks often require:(1)l0 norm or l1 norm;(2)a series of regularization operations used to improve their discriminativeness,and these two operations usually require large computation,especially when the number of training samples is large,the task is particularly huge.But in actual scenarios,timely and efficient classification model is more valuable.In order to solve this problem,Chapter 3 proposes a fast discriminative learning method for fast SAR image classification.This method constructs an effective feature model,with simple l1,? spheres as constraints,instead of computing complex sparse coding;simultaneous learning of two dictionaries,synthetic dictionary and analytical dictionary,to enhance model discrimination;construct nonlinear operators to further mine the nonlinear information of the data.The experimental results on the MSTAR database and the face image data set show that,compared with the same type of method,this method can not only achieve a higher recognition rate,but also a shorter recognition time;2.In order to solve the problem that the existing unsupervised feature selection algorithm cannot effectively mine the structural characteristics of data,Chapter 4 proposes to introduce sparse and low rank constraints in the unsupervised feature selection process;use the anti-interference of l2,1 to outliers to improve the robustness of the model;and unlike the existing algorithm that only constructs a local similarity matrix by kernel function,this method simultaneously learns a similarity matrix during the feature selection process,and uses subspace clustering to mine data the structure of the subspace ultimately guides the selection of features.The experimental results on 7 datasets show that,compared with similar methods,this method can select more discriminative features,which can achieve a higher classification accuracy.3.Low rank representation indicates that the correlation between the data,which is a global constraint,but lacks a description of the local characteristics of the data and cannot effectively discriminate the discriminative information of the data;in addition,the classic subspace learning methods often ignore the data label.In response to these problems,Chapter 5 proposes a block diagonal low-rank representation subspace learning method with local constraints.Construct a block diagonal structure graph to reflect the proximity relationship between data;introduce local constraints in the low-rank representation to mine the local structural characteristics of the data;use the label information of the training samples to guide the construction of the graph,and then use the learned graph to guide the classification of unsupervised data.Experimental results on 6 face data sets show that compared with similar methods,this method has higher discriminativeness and can achieve higher classification accuracy.