SARAH: Semantic-Aware Representation Balance Hashing for Image Retrieval

Deep hashing is vitally important for large-scale image retrieval. Recently, central similarity based deep hashing approaches have shown great advantages for category-level image retrieval; in the existing approaches, however, categories are typically represented by a set of predefined binary vectors which are generated from Hadamard matrix or entry-wisely sampled from Bernoulli distribution. Unfortunately, such kind of category representations lack of discriminativity and semantic information. In this paper, we propose a novel Semantic-Aware Representation bAlance Hashing framework, dubbed SARAH, for category-level image retrieval. Specifically, in SARAH, the category representations are learned to preserve semantic similarities and to maximize pairwise distance; whereas the continuous code of each image is extracted by convolutional network and supervised via a central similarity loss with the corresponding semantic representation which is constructed by the learned category representations. As a consequence, the semantically similar images can be encoded to hash codes with small Hamming distance.