Study On Dstributed Hash Index Technology For Images Retrieval

Along with the development of internet technology and multimedia technology, the amount of the pictures on the internet increase exponentially. How the people find out the pictures that they need is a urgent problem. To solve the problem, the image search systems show up. But the image descriptors that are used in the image search systems have one hundred components at least, even up to hundreds of components. The large amount of descriptors that have many components lead to a tough problem called "curse of dimensionality". To overcome the difficulty, many scholars propose approximate nearest neighbor methods. Locality sensitive hashing is the most effective method of them. In this paper, we make study of a few aspects as follows:1. We propose a novel locality sensitive hashing algorithm called data-dependent locality sensitive hashing on the basis of summarizing the advantages and disadvantages of locality sensitive hashing and its variants. This novel algorithm alleviates the out-of-balance caused by maladjustment of hash functions and index dataset through introducing clustering algorithm. To improve further the search speed, we propose a new search pruning algorithm for data-dependent locality sensitive hashing algorithm. The experimental results show the algorithm can improve the search speed while keeping high precision.2. To realize data-dependent locality sensitive hashing in distributed mode, we must allocate the index dataset between all computing nodes. We must measure the similarity between clusters to realize parallel computation better. We solve the problem of index dataset distribution through introducing the clustering algorithm with restricted items. The experimental results show that the algorithm we propose get better of the algorithm in which the relevance between clusters is ignored.The index algorithm we propose in the paper is constructive for the improvement of performance of image search system.