Cross-Modal Contrastive Representation Learning for Multimedia Retrieval with Noisy Supervision

Cross-modal contrastive representation learning has shown great potential for multimedia retrieval tasks by aligning heterogeneous modalities into a shared embedding space. However, its performance often degrades severely in real-world scenarios where supervision signals are noisy, such as mislabeled cross-modal pairs or ambiguous annotations. To address this challenge, we propose Adaptive Noise-Robust Contrastive Learning (ANRCL), a novel framework designed to enhance cross-modal representation robustness under noisy supervision. Specifically, ANRCL introduces an adaptive noise-robust contrastive loss that jointly exploits cross-modal consistency and intra-modal coherence to dynamically reweight training samples according to their estimated reliability. This mechanism effectively suppresses the influence of noisy pairs while reinforcing the contribution of high-confidence pairs. Experimental results on multiple benchmark datasets demonstrate that ANRCL consistently outperforms state-of-the-art methods in noisy supervision settings, achieving significant improvements in retrieval accuracy and robustness without sacrificing computational efficiency.