Online Matrix Factorization Hashing for Large-Scale Image Retrieval

Abstract

In recent years, approximate nearest neighbor search methods based on hashing have received considerable attention in large-scale data. There are plenty of new algorithms have been created and applied to different applications successfully. However, Due to the coming of big-data era, the data increasing rapidly and constantly. The batch-mode methods cannot process data efficiently. To solve the problem, online hashing has attracted more attention. Online methods can reduce storage and increase speed of computing. But existing online hashing algorithms also have some problems. The first one is the label information often cannot be got. Because of that, supervised approaches are not practicable. Another problem is online hashing methods process data as a stream, so the relations between old data and new arriving data is taken into account. It is the reason why a novel approach is proposed in this paper which combines matrix factorization with the idea of online hashing. This method considers the relationship between the previous data and newly arriving data. In addition, it updates the hashing learning model by the matrix factorization when the new data is arrived. The experimental results demonstrate superiority of the proposed approach. It outperforms most state-of-the-art online hashing methods and batch-mode methods.