Sparse Representations Of Images And Its Application To Image Retrieval

With the developing of the database technology, multimedia technology and network technology, people are increasingly exposed to have a large number of digital image databases. Sparse representation is more and more popular in the field of machine learning and pattern recognition in recent years, because it has important theory value and broad application prospects, sparse representation has been one of the most popular research direction in the field. Sparse representation basic idea is that data signal always has a most sparse representation under a proper sample dictionary in the range of the selected error.This paper proposes two new image retrieval methods:(1) In the CBIR. problem, we propose a new image feature selection algorithm based on group sparse representation, searching image by using the selected feature. First of all, we introduce a regularization-based logistic regression by utilizing both the sparsity and clustering properties. The weight can be effectively worked out by the adaptive spectral gradient algorithm (ANSPG), the most effective features can be selected according to the size of the weights. Finally using the selected effective features for image retrieval on content based retrieval framework.(2) We propose a novel image classification and retrieval algorithm, the basic idea is derived from the sparse representation and kernel. Firstly the training data and test data form their own kernel feature matrix by kernel function. Test image is represented by the linear combination of training images in each of the kernel feature space level. Secondly, group sparse coefficient and the classificatory coefficient can be obtained through multi-task learning in each feature space. Finally, the category of test image can be decided by comparing the residual of reconstruction, and achieving the image retrieval effectively on the basis of classification.