Target Code Guided Binary Hashing Representations with Deep Neural Network

Most existing hashing approaches usually impose some artificial constraints (e.g., uncorrelated and balanced) on hash functions to learn high-quality binary codes, and utilize an optimization strategy which is typically compatible with these hash functions. However, these tight constraints potentially restrict the flexibility of hash functions to fit training data, and result in complicated optimization problem. In this paper, we propose a learning-based hashing method called 'deep supervised hashing with target code'(DSHT) to distill the desirable property in the target coding into hash functions to generate high-quality binary codes. Meanwhile, we incorporate the disparity learning of the intra-class into our proposed method for its generalization. Benefiting from recent advances in deep learning, our framework constructs hash functions as a latent hashing layer in a deep neural network in which binary hashing representations are learned with the guide of target code and semantic information. Experiments on two two large-scale image dataset (MNIST, CIFAR-10) demonstrate that the proposed framework is available, flexible and show comparable performance against other state-of-the-art hashing methods.