Deep Learning of Hash Codes for Content-based Image Retrieval

Abstract

With the explosive growth of online digital visual content, image retrieval becomes popular in our daily life. Hashing methods have been widely and effectively used for large scale image retrieval, which utilizes approximate nearest neighbor concept. Most existing hashing methods encode the input image as a vector of visual features and hash the vector into its binary hash code via projection function or quantization method afterward. This separated two steps make accurate similarity of images lost due to the weak compatibility between a visual feature vector and its binary coding. To avoid this problem, we propose to merge image representation with binary image coding, which maps input image to binary coding directly from a deep learning network. The proposed deep neural network architecture contains two fundamental blocks: the stacked convolution layers of network in network (NIN) with global average pooling for computing effective image representation and the embedded latent layer with bi- nary activation functions for learning binary hash codes. Experimental results show that the proposed method achieved 5-8 percent MAP improvement over the state-of-the-art hashing methods on the CIFAR-10.