Self-supervised Weighted Hashing with Topology Mining
Peishuo Li, Pandeng Li, Wanting Yin · 2023
Recently, although significant progress has been made in unsupervised deep hashing, there are still two neglected but crucial problems. On the one hand, most methods force binary codes to obey the similarity relationships between the original image features, but this noisy supervisory signals can confuse the learning of hash functions. On the other hand, retrieval results of a query often share the same Hamming distance, which gives rise to ambiguous ranking. This paper presents a Self-supervised Weighted Hashing (SWH) framework, which fully explores the intrinsic semantics and local topology structures of representation space to tackle the above problems. First, we mine reliable training guidance from two aspects: high-confidence similarity signals and self-supervised geometric signals. Specifically, by studying the feature similarity statistics, a graph-based sharpening strategy is proposed to disentangle noisy signals and strengthen similarity learning confidence. More importantly, angle prediction and consistency learning based on image rotation transformation are leveraged to capture the geometric semantic information and enhance the robustness of binary codes. Second, a self-supervised learning procedure with a pseudo-siamese structure ( i.e., the hash layer and the weight layer) is proposed to learn weighted binary codes to facilitate adaptive fine-grained distance ranking. Comprehensive experiments on three benchmarks for image retrieval show that SWH outperforms existing methods by 3.5% MAP on average.