Research On Hashing Methods For Content-based Image Retrieval

With the rapid and continuous development of the multimedia technology, network technology and e-commerce, images have brought great convenience to people’s production and life. However, the big image data and high dimensionality make it challenging for image retrieval and data storage. Since text-based image retrieval cannot satisfy the need of people, content-based image retrieval appeared in the late1990s. Searching by image has attracted much attention. For high-dimensional feature data, tree-based methods suffer significantly with their performance typically reducing to exhaustive linear search. Recently, researchers proposed hashing-based methods for image retrieval, which embedding each database item into hamming space. The retrieval system returns the nearest neighbors of a query according hamming distance. Hashing-based methods belong to those for fast approximate nearest neighbor search, which can not only meet users’ retrieval requirement but also have small time and space complexity.Most hashing methods try to learn compact codes with high retrieval performance. This paper makes further research on the problems of existing hash-based methods for content-based image retrieval. The main work and contribution of this paper is listed as follows:First, this paper introduces one kind of feature extraction method based on image content called GIST and several typical hashing methods. We summarize the main ideas and the implementation of these methods. Also we analyze the advantages and disadvantages for improving existing methods.Second, a hashing method based on bilateral random projection is proposed. The approximated low-rank and sparse decomposition of data matrix is studied. Because of the large size of the data matrix, bilateral random projection is used to approximate the low-rank matrix to reduce the time complexity. Projections are obtained after the decomposition by alternating projections on manifolds. Orthogonalizing these projections in batches will produce the final projections for hashing. The variance of hash codes can be reduced by orthogonalizing projections in batches. As a result of the combination of bilateral random projection and batches orthogonalization, the quality of hash codes is improved. Experimental results show that the proposed scheme has improved performance in image retrieval compared with existing methods.Third, an improved algorithm called LPITQ is proposed to enhance the retrieval performance. LPITQ is the combination of ITQ and locality preserving. ITQ can get good hash codes through rotating principle components. Actually, if we have obtained the hashing code of an image, we can improve the code quality through bit selection. In this process, similarity preserving plays the most important role. This enlightened us to put an explicit constraint of locality preserving on ITQ, which leads to more discriminantive and robust hash codes. In the experiments, the improved algorithm performs better than the original scheme.