Research On Algorithms And Applications Of Data Classification And Clustering Based On Sparse Representation

Representation is a fundamental issue in data processing. Finding methods torepresent data effectively and efficiently has been the focus of researchers, and itbecomes especially urgent and important in the current big data age. Sparserepresentation is rooted in neurophysiology, and it becomes quite popular promoted bythe compressed sensing theory proposed in recent years. Sparse representation has beensuccessfully applied in image processing, computer vision and pattern recognition.Focused on data clustering and classification, this thesis attempted to use sparserepresentation to enhance the processing of high dimensional data in unsupervised andsupervised manners. The work and contributions of the thesis include:This thesis systematically reviewed the popular sparse models and the widelyused optimization algorithms solving the sparse regularization problems. Inaddition, the dictionary learning problem and the popular algorithms in theliterature were introduced.Kernel sparse representation was utilized to construct the graph, and a newkind of graph which is called kernel sparse graph was proposed. With thekernel sparse graph, we can model the unsupervised high dimensional data andthen perform spectral embedding and clustering. Experiments show that theproposed kernel sparse graph is superior to existed graphs and the spectralembedding and clustering based on kernel sparse graph is effective.Aiming at data classification, we improved the sparse representation basedclassifier by kernel expansion and structured sparse representation. A novelclassifier called kernel structured sparse representation based classifier wasproposed. Experiments tested on datasets show that the proposed methods canachieve higher accuracy.A novel dictionary learning and sparse representation based unstructured roadsegmentation algorithm was developed. The algorithm uses the local imagepatch as the processing unit; a dictionary was learned and it can be updated inreal time by online dictionary learning. The local patches of the test image canbe classified by the reconstruction errors of sparse representation on the learneddictionary. Extensive experiments show that the proposed algorithm is suitablefor various unstructured environments and is robust to illumination, shadowand water stains. This algorithm achieved good results in visual navigation of acertain type of autonomous land vehicle in unstructured environment.