Deep semantic preserving hashing for large scale image retrieval

Hashing approaches have got a great attention because of its efficient performance for large-scale images. This paper, aims to propose a deep hashing method which can combines stacked convolutional autoencoder with hashing learning, where the input image hierarchically maps to the low dimensional space. The proposed method DCAH contains encoder-decoder, and supervisory sub-network, that generates a low dimensional binary code in a layer-wised manner of the deep conventional neural network. To optimizing the hash algorithm, we added some extra relaxations constraint to the objective function. In our extensive experiments on ultra-high dimensional image datasets, our results demonstrate that the decoder structure can improve the hashing method to preserve the similarities in hashing codes; also, DCAH achieves the best performance comparing to other states of the art approaches.