Research On Image Reranking Based On Image Hashing

With the rapid development of the internet and multimedia technology, image has become a significant form of multimedia information. More and more applications lead to the emergence of Content-Based Image Retrieval. In order to adapt to the retrieval requirements of large-scale datasets and reduce the influence of "dimension disaster", hashing-based image retrieval has been proposed. Image hashing is encoding the visual features into binary hash codes, and Hamming distance is adopted to measure the similarity between the hash codes. The main advantages of hashing-based image retrieval are summarized as follows. First, the computation speed of Hamming distance between hash codes is fast enough to deal with the large-scale datasets. Second, using hash codes to represent images can largely save the memory space. However, there is limitation of hashing-based image retrieval. Because the Hamming distance value is an integer, there will be hundreds or thousands of images sharing the same Hamming distance to the query image on large-scale datasets. How to rank images with the same Hamming distance is no longer a negligible problem. This paper studies image hashing methods and hashing-based image reranking methods. The main contributions are as follows:(1) This paper introduces several typical hashing methods, including random projection based hashing methods and PCA projection based hashing methods. We analyze the pros and cons of various hashing methods from theoretical analysis and experimental validation.(2) We improve the PCA projection based reranking method. First, we introduce and realize the compact hash codes based reranking method (QAIS) and PCA projection based reranking method (QsRank), and analyze their advantages and disadvantages. The theoretical foundation of QsRank reranking method is that ε neighborhood relationship is preserved through projection, and therefore, it fails to rerank random projection based hashing methods. Moreover, QsRank reranking method does not depend on Hamming distance to rank images, so it decelerates the retrieval speed. Combining high efficiency of hashing methods and advantages of QsRank method, this paper proposes IQsRank method. The experimental result has demonstrated IQsRank method has higher precision rate and shows robustness to different hashing methods.(3) This paper proposes a query-adaptive reranking method (QAR) for hashing-based image retrieval. To solve the problem of current hashing-based reranking methods, this paper focuses the characteristics of projection function. With the combination of semantic information of datasets and projection function, category-specific weights are learnt offline. Based on hashing retrieval results, query-adaptive weights can be computed and adaptive Hamming distance can be constructed. The adaptive Hamming distance makes the discrete Hamming distance value continuous, solving the ranking problem of images with the same distance. The experimental results have proved the effectiveness of proposed method.This paper is supported by the Major Projects of National Natural Science Foundation of China (Grant No.70890083).