Research On Large-scale Image Retrieval Technology Using Sparse Representation

Image retrieval is an important part of the information retrieval technology. To acertain extent, its development solves the semantic gap between the image and humanunderstanding. In the face of large-scale image data, to provide efficient image retrievaltechnology has become the research hotspot in the field of computer.The thesis focuses on image feature extraction, feature fusion, image representation,classification and large-scale data processing in image retrieval technology, and putforward a number of algorithms and technical schemes. Main research works are asfollows:(1) It discussed the description and matching of image features in image retrieval.Improved the dominant color algorithm based on the relative error distance. Experimentalresults showed that the colors match algorithm has good effects using the colornormalization and block weighted matching method.(2) It put forward the texture representation method based on nonsubsampledcontourlet and gray level co-occurrence matrix feature fusion. The method could moreaccurately represent image spatial domain and frequency domain multiple information.(3) Introduced sparse representation method; built the image classification methodbased on sparse representation. The method used rough classification based on detectedimage to narrow the scope of image retrieval. The method mainly used the unique nature ofsparse representation to train the sparse dictionary, and determined target category throughthe sparsity and residuals.(4) In the face of large-scale image retrieval efficiency problem, probed a retrievalmethod based on the hierarchical K-means clustering algorithm to reduce the problem size.In the cluster and the training process, we introduced parallel computing method, designedparallel computing method of K-means to shorten clustering consuming time.In this thesis, the proposed method is demonstrated, and results show that the proposedscheme is effective.