| With the development of digital technology,the reduction of hardware cost,the popularization of intelligent devices and the evolution of Web,more and more images are created and netizens can share their images on the Internet as they wish.It has become a great challenge to retrieve images on the Web accurately,quickly and efficiently.So far,there has already exist some image retrieval methods,i.e.,text-based image retrieval,TBIR for short,content-based image retrieval,CBIR for short.However,the existing methods cannot handle unannotated image retrieval with free words well,i.e.,TBIR is limited to a given list of labels,CBIR requires a seed image,which is hard to obtain in some cases.So this thesis focuses on the problem and following are the main contributions:Firstly,due to the challenge of retrieving unannotated images with high-level concepts,we propose a novel method based on concept decomposition.With the help of Wikipedia,we can get a list of proper sub-concepts with concrete visual representation.As a result,links between text space and visual space are built properly and the semantic gap between low-level visual features high-level semantics is reduced.Experiment result shows the effectiveness of our method.Secondly,due to the unsatisfactory performance of retrieval of Web images and their surrounding free text with keywords,we propose a new framework,named as MEIR,which applies multi-modal enhancement to Web images and their surrounding free text and improves their similarity measurement.Then,we can enrich the keywords and implement text complement on the Web images if they lack of text.Experiment result shows that multi-modal enhancement is helpful for Web image retrieval. |
