Efficient Incremental Near Duplicate Detection Based on Locality Sensitive Hashing

In this paper, we study the problem of detecting near duplicates for high dimensional data points in an incremental manner. For example, for an image sharing website, it would be a desirable feature if near-duplicates can be detected whenever a user uploads a new image into the website so that the user can take some action such as stopping the upload or reporting an illegal copy. Specifically, whenever a new point arrives, our goal is to find all points within an existing point set that are close to the new point based on a given distance function and a distance threshold before the new point is inserted into the data set. Based on a well-known indexing technique, Locality Sensitive Hashing, we propose a new approach which clearly speeds up the running time of LSH indexing while using only a small amount of extra space. The idea is to store a small fraction of near duplicate pairs within the existing point set which are found when they are inserted into the data set, and use them to prune LSH candidate sets for the newly arrived point. Extensive experiments based on three real-world data sets show that our method consistently outperforms the original LSH approach: to reach the same query response time, our method needs significantly less memory than the original LSH approach. Meanwhile, the LSH theoretical guarantee on the quality of the search result is preserved by our approach. Furthermore, it is easy to implement our approach based on LSH.