面向Web图像检索的基于语义迁移的无监督深度哈希  

Unsupervised deep hashing based on semantic transfer for Web image retrieval

作　　者：许胜 陈盛双[1] 谢良 Xu Sheng;Chen Shengshuang;Xie Liang(School of Science,Wuhan University of Technology,Wuhan 430070,China)

机构地区：[1]武汉理工大学理学院

出　　处：《计算机应用研究》2019年第9期2866-2871,共6页Application Research of Computers

基　　金：国家自然科学基金青年基金资助项目(61702388)

摘　　要：当前主流的Web图像检索方法仅考虑了视觉特征,没有充分利用Web图像附带的文本信息,并忽略了相关文本中涉及的有价值的语义,从而导致其图像表达能力不强。针对这一问题,提出了一种新的无监督图像哈希方法--基于语义迁移的深度图像哈希(semantic transfer deep visual hashing,STDVH)。该方法首先利用谱聚类挖掘训练文本的语义信息;然后构建深度卷积神经网络将文本语义信息迁移到图像哈希码的学习中;最后在统一框架中训练得到图像的哈希码和哈希函数,在低维汉明空间中完成对大规模Web图像数据的有效检索。通过在Wiki和MIR Flickr这两个公开的Web图像集上进行实验,证明了该方法相比其他先进的哈希算法的优越性。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.

关 键 词：语义迁移 图像哈希 WEB图像检索 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]