Research On Weighted Representation-based Classification

With the rapid development of the Internet,the field of pattern recognition has become a hot research object in this era of artificial intelligence.Representation-based classification?RBC?has become a promising pattern recognition classification method in recent years.Among them,sparse representation is widely favored by scholars for its unique sparseness,and collaborative representation is reduced because of its closed form solution.The amount of calculation caused widespread concern.In this thesis,the two are combined with each other and their advantages are complementary.At the same time,locality constraints are considered,and adaptive weighting is introduced to improve the classification performance of pattern recognition.This thesis conducts the following research work on the improved weighted representation classification method:?1?Considering the locality of data in LRM,the weighted two-phase linear reconstruction measure-based classification method?WTPLRMC?is proposed.In WTPLRMC,the first phase determines the representative training samples from all training samples through LRM,and the second phase uses the locality of the data to constrain the linear reconstruction coefficients of the representative training samples selected in the first phase.Similarity weights between samples reflect each test sample and representative training samples.?2?Representation-based classification?RBC?can be used for classification well as the representation coefficient of the linear reconstruction measure?LRM?.Therefore,two classification methods based on the sparsity augmented collaborative representation-based classification method?SA-CRC?are proposed.The one is the weighted enhancement linear reconstruction measure-based classification method?WELRMC?that introduces the localities of data into SA-CRC.The other one is the two-phase weighted enhancement linear reconstruction measure-based classification method?TPWELRMC?that integrates both the coarse to fine representation and SA-CRC,and consider the local advantages between training and testing samples.?3?Although most of the CRC methods seem to be discriminative for classification,they fail to discover the prominent discrimination among different classes and the dominant representation of the correct class for each testing samples simultaneously in the representation-based feature space.For the classification based on the ideal representation,the pattern distinguishes the representation of a specific category between different categories,and the true category of each test sample should dominate the coding.To solve this issue,the class mean-based weighted discriminative collaborative representation method?CMWDCRC?is proposed.In the proposed CMWDCRC model,one discriminative term is the regularization of pairs of any two class-specific representations,which focuses on the discrimination among class-specific representations from different classes.And the other newly designed discriminative term,which aims to competitively improve the class-specific representations to achieve the dominant representation from correct class,is the constraint of weighted distances between the class-specific representations and their corresponding class mean vectors.As a result,the true class of the testing sample has more contribution to coding and the other classes have less contribution.To further enhance the robustness of CRC in the noise scenarios of image classification,the designed CMWDCRC with7)₂-norm representation fidelity of coding residual is extended to the robust CMWDCRC?R-CMWDCRC?with7)₁-norm representation fidelity.?4?Design and implement an image recognition system based on the weighted representation classification method.The superiority of the proposed algorithm is shown to the reader simply and intuitively through the image recognition system,which is convenient for the reader to observe the contrast effect between the improved algorithm and other RBC methods.