Image Retrieval with Well-Separated Semantic Hash Centers

Liangdao Wang, Yan Pan, Hanjiang Lai, Jian Yin; Proceedings of the Asian Conference on Computer Vision (ACCV), 2022, pp. 978-994

Recently, some point-wise hash learning methods such as CSQ [1] and DPN [2] adapted "hash centers" as the global similarity label for each category and force the hash codes of the images with the same category to get closed to their corresponding hash centers. Although they outperformed other pairwise/triplet hashing methods, they assign hash centers to each class randomly and result in a sub-optimal performance because of ignoring the semantic relationship between categories, which means that they ignore the fact that the Hamming distance between the hash centers corresponding to two semantically similar classes should be smaller than the Hamming distance between the hash centers corresponding to two semantically dissimilar classes. To solve the above problem and generate well-separated and semantic hash centers, in this paper, we propose an optimization approach which aims at generating hash centers not only with semantic category information but also distinguished from each other. Specifically, we adopt the weight of last fully-connected layer in ResNet-50 model as category features to help inject semantic information into the generation of hash centers and try to maximize the expectation of the Hamming distance between each two hash centers. With the hash centers corresponding to each image category, we propose two effective loss functions to learn deep hashing function. Importantly, extensive experiments show that our proposed hash centers and training method outperform the state-of-the-art hash models on three image retrieval datasets.