Online weighted hashing for cross-modal retrieval

Zining Jiang, Zhenyu Weng, Runhao Li, Huiping Zhuang, Zhiping Lin

Abstract: In recent times, there has been a growing interest in online hashing for cross-modal retrieval tasks with streaming data. This technique involves encoding streaming data into compact binary codes, which has the potential to reduce data storage requirements and enhance search efficiency. However, it is important to note that as the length of binary codes shortens, the retrieval performance deteriorates rapidly due to an ambiguity issue caused by Hamming distance. To solve this problem, we propose a novel online hashing method, Online Weighted Hashing (OWH), to improve the retrieval performance for short binary codes by learning weights online to replace Hamming distance with weighted Hamming distance. By learning different weights on each bit of binary hash codes, it solves the ambiguity issue associated with Hamming distance as well as preserves more semantic information, and thus improves the accuracy of cross-modal retrieval. The online weight learning problem is formulated by exploiting the similarity between newly coming data and existing binary codes, and an efficient optimization algorithm is designed to solve this problem. Extensive experimental results on three benchmark datasets demonstrate that the proposed method outperforms state-of-the-art online and offline cross-modal hashing methods in terms of retrieval accuracy.