Label-Corrected Weighted Multi-Similarity Loss for noisy cross-modal retrieval

Label noise causes semantic inconsistency and training sample pair imbalance, severely degrading the performance of noisy cross-modal hashing in real-world retrieval scenarios. To address this problem, this paper proposes Label-Corrected Weighted Multi-Similarity Loss (LC-WMSL), a unified framework integrating adaptive label correction and cross-modal embedding optimization for noisy cross-modal retrieval. LC-WMSL includes three core components: a label dependency-aware correction module that models instance-label and label-label correlations via dynamic graph learning and leverages deep network memorization for adaptive denoising; a multi-similarity weighted hashing module that mitigates pair redundancy and imbalance through joint optimization of intra-sample self-similarity and inter-sample relative similarity; and a label-guided network for fine-grained cross-modal semantic alignment. Extensive experiments on MIRFlickr25K, MS-COCO and NUS-WIDE show LC-WMSL outperforms state-of-the-art methods by 2.13%, 1.97% and 4.61% average mAP on image-to-text and text-to-image tasks, maintaining robust retrieval performance even at 80% label noise. Ablation studies further validate the independent effectiveness of each core component in boosting hash code discriminability and enhancing overall model generalization. The source codes of LC-WMSL are downloaded from https://github.com/XS185/LC-WMSL . • We propose LC-WMSL, a noise-robust cross-modal hashing framework. • We design a semantic dependency-based label correction mechanism. • We establish a multi-similarity weighted hashing optimization paradigm. • We validate the superiority via extensive experiments and ablation studies.