Binary code learning via optimal class representations

Hashing is an attracting technique for fast retrieval due to its low storage and computation costs. By hashing, each high-dimensional vector is mapped into a low-dimensional binary code vector and retrieval is performed in the Hamming space. Recently several hashing methods have been proposed, among which, supervised hashing methods have shown great performance by incorporating the supervision information. However, most previous supervised methods simply focused on the pairwise label information of data, and ignored the structure information and relationship within data. To tackle this problem, we propose to learn binary codes by explicitly taking into account class semantic relatedness. Specifically, a set of binary codes is computed according to the intrinsic class similarities in data and serves as the optimal class representations. We show that, by mapping images onto the optimal representation of their corresponding classes, our proposed method outperforms several other state-of-the-art supervised hashing methods in image retrieval on three large-scale datasets.