Multi-view Subspace Learning Algorithm And Its Application In Classification And Recognition

Multi-view data represent the different characteristics of objects from different views.How to make full use of the common feature information of these views and the unique feature information of each view to avoid the limitations of a single view,so as to comprehensively and profoundly understand the target object has become a hot research issue in the academic field.This kind of algorithm is called multi-view subspace learning algorithm.This kind of algorithm has been widely concerned and applied in computer vision,aerospace and other fields,so it is of great significance and value to study it in depth.Despite the rapid development of multi-view subspace learning algorithm in face recognition and object classification,it still faces many challenges,such as large data differences and unknown view information.In this thesis,aiming at the challenges such as the large difference of multi-view data and the unknown view information of the test set in the classification and recognition task,the feature extraction part of the traditional framework is improved by using the complementary and proprietary characteristics of multi-view data,and the algorithm framework of multi-view learning algorithm applied in the classification and recognition task is designed.On this basis,two algorithms are proposed,which are low rank joint sparse representation algorithm based on common subspace and generalized low rank joint sparse representation algorithm based on common subspace learning.The two algorithms establish the relationship between the multi-view data through the common subspace,which can effectively reduce the difference between the data views and deal with the application scenarios with unknown test sample view information.And using the complementary characteristics of multi-view data,when some of the view data are affected by noise points,the data information of other views can bring corresponding supplement,and the algorithm can effectively deal with the application scenarios of data pollution and destruction.For the extracted features,both algorithms impose low rank constraints to expose the global structure of multi-view data and align the data,and two joint sparse constraints are respectively applied to fuse the extracted features and reduce the redundant information of the dictionary.Through these ways to improve the discrimination of the algorithm.According to the proposed algorithms,the optimization formula is given.And based on ADMM,CCCP and Lagrange dual method,the reasonable solution scheme is designed to effectively solve the two algorithms.At the same time,the low rank problem in the optimization formula of the algorithm is accurately solved by an exponential approximation.Finally,this thesis designs a reasonable experiment,using CMU-PIE,CASIA-HFB and other multi-view datasets to evaluate the proposed algorithm,and compared with seven mainstream multi-view learning algorithms.The experimental results fully show that the two proposed multi-view subspace learning algorithms have good recognition performance,can effectively solve the challenges in classification and recognition tasks,have good robustness,and effectively achieve the research objectives of this thesis.The algorithms proposed in this thesis have important research value and significance in theoretical innovation and practical application.