Deep regional detail-aware hashing

Deep hashing has been widely studied in recent years. However, most existing methods learn hash codes from the whole image ignoring the region details which are important for representing the precise semantic contents. To tackle this issue, we proposed a deep regional detail-aware hashing (DDAH) to fully utilize this detail information. Specifically, to well reflect the influence of details on the hash codes, we handle these details in a “near-Hamming” space instead of directly fusing them in the original image feature space. Furthermore, based on the framework of DDAH, we can capture the details in the network without the need to partition the original image to different regions manually. To be more specific, considering multiple regions as overlapping subimages, we first design a deep network to learn multiple regional details from these subimages, and then fuse them in the near-Hamming space, which is highly related to Hamming space (i.e., hash code space). Finally, these regional details in the near-Hamming space are directly used to generate hash code of the corresponding image. In addition, a self-similarity loss term is proposed to force these regional details together in the near-Hamming space. In brief, compared to existing hashing methods, the proposed DDAH not only utilizes detail information for hash learning, but also incorporates them into the final hash codes. Extensive experiments on three datasets have indicated that DDAH outperforms most existing models, verifying its effectiveness.