Triple collaborative asymmetric deep hashing for multi-label image retrieval

Abstract Due to its efficient storage and fast retrieval capabilities, hashing methods have garnered significant attention in image retrieval applications. However, most existing methods rely on semantic similarity matrices to indicate whether two images are similar or not, which neglects the guiding role of semantic signals in the multimodal information flow. Moreover, conventional symmetric hashing architectures encode both query and database images simultaneously, which increases the time overhead of the model. To address aforementioned limitations, we propose a Triple Collaborative Asymmetric Deep Hashing (TCADH) framework for multi-label image retrieval, which maps two asymmetric image streams and one semantic label stream into a common discrete Hamming space, achieving collaborative alignment of multi-source information. Specifically, we first build a Siamese network based on ResNet34 to extract features from both original and augmented images, while incorporating a sign-consistency-guided feature enhancement block into the backbone to mitigate information loss caused by feature dimensionality reduction. Furthermore, we introduce a semantic label stream to capture fine-grained semantic supervision and employ a weighted pairwise loss to maximize semantic consistency among different data streams from the same image source, thereby achieving semantic alignment among original images, semantic labels, and augmented images in the Hamming space. Finally, we adopt an asymmetric learning strategy to capture similarities between hash codes and real-valued image features, improving retrieval precision and accelerating model convergence. Extensive experiments are conducted on four multi-label datasets: NUS-WIDE, MIRFLICKR-25K, MS-COCO, and VOC2012, the results demonstrate that the proposed TCADH outperforms state-of-the-art deep supervised hashing methods in multi-label image retrieval tasks.