Fast and Efficient Image Retrieval via Fully-Convolutional Hashing Network

With the rapid development of information technology, content-based image retrieval and related technologies have become increasingly important. The hash method can represent an image with a sequence of binary codes consisting of 0 and 1, while the application of a convolutional neural networks can learn directly from the image to a discriminative binary code. We propose a deep hash algorithm based on full convolutional neural network, which can reduce the number of network parameters and have the retrieval performance not lower than the cutting-edge method in this field. The proposed network structure uses convolutional layers and the global average pooling layer to replace fully-connected layers in current deep hashing network structures, which significantly reduces the complexity of the network and improved its training performance. This method is called Fully-convolutional hashing networks (FCHN), and experiments were carried out on several publicized datasets to verify the effectiveness of the method.