A Label Noise Robust Cross-Modal Hashing Approach

Abstract

Cross-modal hashing has attracted more and more research interest for its high speed and low storage cost in solving cross-modal approximate nearest neighbor search problem. With the rapid growth of social networks, a large amount of information is generated every day, which inevitably contains some noisy information. However, most existing cross-modal hashing methods do not take label noise into consideration and simply assume that all training data are completely reliable. Therefore, these methods will be affected by label noise and reduce the effectiveness on the real-world data. In this paper, we propose a novel end-to-end cross-modal hashing method called label noise robust cross-modal hashing (LNRCMH) to solve the cross-modal hashing problem on data with label noise. LNRCMH first calculates the local outlier factor (LOF) for each instance to evaluate the probability that the instance is corrupted by the label noise. Then LNRCMH assigns lower weights to the instances with high probabilities to be corrupted. Finally, LNRCMH uses different neural networks to learn features for instances from different modalities and transforms them into binary hash codes. Our experimental results on multi-modal benchmark datasets demonstrate that LNRCMH performs significantly better than other related and comparable methods with noisy label annotations. Our approach also achieves competitive results in noise-free situations.