Proxy-Guided Semi-supervised Hashing for Cross-modal Retrieval

With the exponential growth of multimodal data, cross-modal retrieval has become increasingly important. While supervised deep hashing methods have achieved state-of-the-art performance, their heavy reliance on large-scale annotated data limits their practical application. Semi-supervised methods aim to mitigate this by leveraging unlabeled data but often struggle with unreliable pseudo-labels and inadequate semantic structure modeling. To address these challenges, this paper proposes a novel Proxy-Guided Semi-Supervised Hashing (PGSSH) framework. Inspired by the success of proxy-based mechanisms in supervised methods, PGSSH introduces a semantic proxy matrix to act as stable anchors for both labeled and unlabeled data, explicitly modeling category structures in the Hamming space. Furthermore, we design a Confidence-Aware Pseudo Label Generator (CAPLG) that robustly estimates pseudo-label quality by integrating proxy similarity, cross-modal consistency, and an adaptive threshold. A comprehensive Proxy-Guided Semantic-Preserving (PGSP) Loss is proposed to jointly optimize supervised learning, confidence-aware pseudo-labeling, and cross-modal ranking. Extensive experiments on three benchmark datasets (MIRFLICKR-25K, NUS-WIDE, and MS COCO) demonstrate that PGSSH significantly outperforms state-of-the-art semi-supervised methods.