RSPH: Robust self-paced hashing for cross-modal retrieval

Recently, hashing technology has attracted considerable attention in cross-modal retrieval due to its high retrieval efficiency and low storage overhead. Therefore, many cross-modal hashing (CMH) methods have been proposed and performed well. However, most existing CMH methods treat all samples equally without discrimination, which may lead to suboptimal local minima. Besides, most existing approaches usually leverage modality-specific information when learning hashing functions, which cannot effectively capture both intra-and inter-modality correlations in multimodal data. To mitigate these problems, a novel Robust Self-Paced hashing (RSPH) approach is developed for cross-modal retrieval. Concretely, in the proposed RSPH, the multimodal data are progressively integrated into the hashing learning process from easy to hard, helping to avoid poor local minima. Besides, the ℓ _2,1-norm is employed in the proposed method to further reduce the impact of outliers and noise. To fully exploit multimodal information, this method introduces a collaboration projection scheme that captures both modality-specific properties and cross-modal correlations, resulting in more effective hash functions. Experimental results on three benchmark datasets validate our RSPH can achieve state-of-the-art performance, demonstrating its efficacy.