基于大批量训练和正交正则化的跨模态哈希方法

he cross-modal hashing methods based on deep learning use the small batch training method to train their model. However, it cannot get a good gradient using this training method due to the limited number of samples in each parameter update, which affects the retrieval performance of the final trained model. To solve the problem, this paper proposed a new cross-modal hashing, which used large batch training and introduced orthogonal regularization to increase the stability of this kind of training. And to consider the discreteness of hash codes, the objective function added the distance between hash codes and features which made hash codes to represent data more realistically. Extensive experiments on two widely used public datasets in cross-modal hashing show that this method achieves better performance than several existing hashing methods.