Research On Theories And Applications Of Similar Image Retrieval System Based On Improved Multi-index Hashing

In recent years, with the development of Internet technology, multimedia data grows rapidly. A huge amount of multimedia data has no use unless we can access and retrieve it efficiently. In some scenarios, such as similar image retrieval,songs retrieval according to sound clips etc, traditional text-based information retrieval technology can’t satisfy people’s needs. There are some disadvantages of traditional text-based information retrieval technology. For examples, the ability of text-based expression is limit in those scenarios and the cost of labor is heavy when handling large scale multimedia data retrieval. Therefore, information retrieval on multimedia data like image, sound etc which is more intuitive than text becomes more and more important.The contribution of this paper is to design and implement large scale similar image retrieval system based on improved multi-index hashing. The core algorithm of the system is multi-index hashing. we apply this algorithm to k-nearest neighbor search on feature vectors of images. In application, we spot that the algorithm will have poor performance under certain conditions. In order to solve the problem, we optimize the hash strategy of multi-index hashing. Before building hash tables, we will rearrange the position of elements of the feature vectors.We can solve the problem to some extent after improving the hash strategy of multi-index hashing.The similar image retrieval system can utilize the computing power of multicores in modern CPU to increase the throughput. However, instead of parallelizing multi-index hashing, the system will process requests in multi-threads and execute mult-index hashing serially in per thread due to the difficulty of parallelizing multi-index hashing. To support the mechanism, we design a lock-free queue to decrease the cost of synchronizing the shared data. In experiment, we can show that the lock-free queue has significant performance when comparing to the queue based on mutex lock.In the last of the paper, we will show the basic function of the similar image retrieval system. Although the system is a demo currently, the retrieved results satisfy our needs and the speed is very fast.