Topics Oriented Image Annotation And Retrieval

With the development of social informatization and the popularization of portable cameras,on one hand,people,as creators,can share pictures anytime and anywhere,which forms massive images on the Internet;on the other hand,as users,people need to find satisfying pictures from massive resources.That brings challenges to the image management,organization and retrieval,so the image retrieval research has increased.So far,the research of image retrieval has achieved remarkable results.But from the perspective of subject distribution,this research is mainly conducted by the Computer Science which focuses on the innovation of the retrieval algorithm development;however,there is only a little research on image retrieval from the Library and Information Science field which mainly focuses on the theoretical aspect.In this paper,the author will not only be concerned about the image retrieval technology,but also consider the discipline characteristics of our field,and make some new attempts in the image retrieval research,such as presenting the image retrieval results with images and labels.Firstly,this paper investigated the domestic and abroad existing research literature and summarizes the advantages and disadvantages of the existing image retrieval techniques and image semantic annotation methods.Secondly,choose topics of high concern in daily life(e.g.haze,waterlogging,flowers and maple leaves),then collect each topic's images with text descriptions from the Internet.Then,process the text and image information respectively,which includes three aspects:(1)Process the text information to form the image semantic annotation,and mainly involves the keyword extraction and the semantic extension.(2)Extract image features to form the visual feature description,and mainly includes the extraction process of improved color histogram,Local Binary Pattern(LBP)and grayscale co-occurrence matrix.This part mainly describes the improved algorithm which improves the traditional color histogram by weighting the blocked images with color entropy.(3)Use all visual features to construct multi classifier model based on SVM,then apply the image classification model to content based image retrieval,which can establish the connection between visual features and semantic classes,so then improve retrieval performance by narrowing the semantic gap.Finally,the image retrieval system is designed and implemented based on the image annotation.The system supports several retrieval methods:text based image retrieval(using keywords to search images),content-based image retrieval(using an image to search images)and these two methods combined.Moreover,this system displays more fine-grained results with images and text labels in order to improve user experience.In addition,the feasibility of the system and the effectiveness of the proposed method are verified by the related experiments.