Search

Most existing Web image retrieval approaches only consider visual features. They ignore the valuable semantics involved in the associated texts,and fail to take advantages of text. This paper proposed a new unsupervised visual hashing approach called STDVH. Firstly, it extracted the semantic information of the training text by spectral clustering. Then, it constructed a deep convolutional neural network to transfer the text semantic information into the learning of the image hash code. At last, it trained the image hash codes and hash functions in a unified framework, and completed the effective retrieval of largescale image data in low-dimensional Hamming space. Experiments on two publicly available image datasets Wiki and MIR Flickr indicate that the proposed approach can achieve superior performance over other state-of-the-art techniques. [ABSTRACT FROM AUTHOR]