Co-sparsity regularized deep hashing for image instance retrieval

In this work, we tackle the problem of image instance retrieval with binary descriptors hashed from high-dimensional image representations. We present three main contributions: First, we propose Co-sparsity Regularized Hashing (CRH) to explicitly optimize the distribution of generated binary hash codes, which is formulated by adding a co-sparsity regularization term into the Restricted Boltzmann Machines (RBM) based hashing model. CRH is capable of balancing the variance of hash codes per image as well as the variance of each hash bit across images, resulting in maximum discriminability of hash codes that can effectively distinguish images at very low rates (down to 64 bits). Second, we extend the CRH into deep network structure by stacking multiple co-sparsity constrained RBMs, leading to further performance improvement. Finally, through a rigorous evaluation, we show that our model outperforms state-of-the-art at low rates (from 64 to 256 bits) across various datasets, regardless of the type of image representations used.