Semi-supervised Dictionary Learning Based On Atom Graph Regularization

With the development of technologies in the artificial intelligence,dictionary learning has been a hot research in the machine learning and pattern recognition field.Traditionally,it can be divided into two main groups: the supervised dictionary learning and semi-supervised dictionary learning.To the former,the classification results rely on the number of labeled samples and the latter doesn't make full use of the latent manifold structural information among the training samples,and thus,the classification accuracy can not reach the best performance.To solve the above problems,this master's thesis mainly completed three projects:1.In the framework of semi-supervised learning,the graph laplacian defined on the atoms and the corresponding sparse codings is embedded into the framework of the dictionary learning,so a semi-supervised dictionary learning based on atom graph regularization(AGR-SSDL)was proposed.2.Basing on the above model,a feedback path which can incorporate timely the information from the estimated labels of unlabeled samples into the dictionary learning process was added,so the newlyproposed work is called as self-taught semi-supervised dictionary learning based on the atom graph regularization(ST-SSDL).3.Basing on the above work,a non-negative and self-taught semi-supervised dictionary learning based on atom graph regularization(NNST-SSDL)was formulated by imposing the non-negative constraint on sparse coefficient matrix.Extensive experiments on several benchmark databases demonstrate the superior performance of the proposed method compared with the state-of-the-art semi-supervised dictionary learning methods.