Font Size: a A A

Research On Classification Algorithms Based On Multi-layer Dictionary Learning

Posted on:2020-05-14Degree:MasterType:Thesis
Country:ChinaCandidate:J ZhaoFull Text:PDF
GTID:2428330602951047Subject:Engineering
Abstract/Summary:PDF Full Text Request
Dictionary learning is an important research direction in the field of machine learning and pattern recognition.It is also an important method in the field of representation learning.It mainly learns the dictionary atoms so that the data can be represented by these atoms with great sparsity.The methods based on dictionary learning for classification and recognition have achieved good results in many applications.Based on the multi-layer dictionary learning framework,this paper mainly studies how to enhance the discriminative ability of multi-layer dictionary learning methods and apply them to practical problems.The research results obtained are as follows:Firstly we point out that,for the conventional Multi-Layer Dictionary Learning(MDL)method,the label information of the data is not utilized so that the learned representation coefficients is not discriminative.To solve this issue,we propose a Discriminative Multi-Layer Dictionary Learning based on Regression Error(REDMDL).By introducing a regression error term,RE-DMDL can enhance the discriminative ability of the representation coefficients while considering the reconstruction error,and further increase the “label distance” by utilizing a nonnegative slack variable.Thus,Re-DMDL is more suitable for the classification task and can improve the accuracy of MDL methods.At the same time,the constraint of the representation coefficient is relaxed to the L2 norm,which reduces the time overhead for training the algorithm.The experimental results verify that the proposed algorithm,RE-DMDL,can improve the classification accuracy of MDL method on classification tasks.Secondly,this paper considers how to obtain more discriminative representation coefficients of MDL;from the perspective of graph constraint,a Multi-Layer Dictionary Learning based on Label Graph Constraints(LGC-MDL)method is then presented.By constructing a label graph,the distance between the representation coefficients of the same class of samples is minimized and the distance between different classes of samples is maximized.Thereby the discriminability of the representation coefficients is enhanced while the reconstruction error is optimized.The experimental results verify that the proposed algorithm,LGC-MDL,can further improve the classification accuracy of MDL related methods on classification tasks.Thirdly,we study the Projection Twin Support Vector Machine(PTSVM)and find that it is sensitive to noises and outliers and it does not consider the local manifold structure of the data.To this end,a Fuzzy Projection Twin Support Vector Machine based on Global and Local Information(FGLPTSVM)is proposed.By borrowing the idea of fuzzy membership,each sample point is assigned a fuzzy membership degree according to its confidence level belonging to a certain class,which reduces the influence of noises and outlier points in the model.At the same time,by constructing a global and local graph constraint,the local manifold information of the data is well maintained.The experimental results prove that the proposed algorithm has better classification accuracy compared to traditional methods,especially when noises and outliers exist.Finally,it is realized that radar emitter identification is a key problem to be solved by many practical systems such as enemy identification and threat warning.The REDMDL algorithm and LGC-MDL algorithm proposed in this paper are applied to the individual identification of radar emitters.The experimental results proves that the two algorithms can improve the classification accuracy of the radar emitter identification system with great value in real engineering application.
Keywords/Search Tags:Dictionary Learning, Multi-Layer Dictionary Learning, Classification Algorithm, Support Vector Machine, Radar Emitter Identification
PDF Full Text Request
Related items