An Efficient Method For Content-and-Text Based Image Retrieval

With the rapid development of information technology and social network,the increasing scale of image databases brings new challenges for image retrieval technology.It is an urgent problem to propose efficient retrieval method for large scale image database in the information retrieval field.Traditional image retrieval technology can be classified into two major categories:text-based image retrieval(TBIR)and content-based image retrieval(CBIR).TBIR searches similar images by matching the query keywords with the textual annotations of images.It is simple and fast,but dependent on manual labelling,which brings some drawbacks such as expensive labor and time cost,subjectivity and polysemy of textual description and so on.CBIR measures similarity of images with low-level visual features extracted from images,which avoids the disadvantages of manual labelling.However,there is a bottleneck problem in CBIR:the semantic gap between low-level visual features and high-level semantics.To bridge the semantic gap of traditional methods,fusion of textual and visual information has become a research hotpot of image retrieval recently.Nevertheless,existing studies in this field pay major attention to feature model,hashing method and various fusion techniques to enhance the retrieval quality and the speed of image processing,but few works concentrate on how to improve the retrieval efficiency by index mechanism.Therefore,in this paper,we focus on improving the retrieval efficiency for hybrid image retrieval based on fusion of content and text.First,adopting linear fusion technique based on hybrid similarity of visual features and textual annotations,we transform the top-k image retrieval problem into a problem similar to the spatial keyword query.Then,we designs a novel hybrid index called HIR-tree by combing Manhattan Hashing,Inverted Index and M-tree,and introduces a set of important metrics for HIR-tree.Meanwhile,we propose an efficient retrieval solution of top-k hybrid image retrieval based on them,including three phases: preprocessing,construction algorithm and top-k query algorithm of HIR-tree.At last,rigorous theoretical analysis is present for our method,and a series of experiments based on benchmark image databases show that our method can remarkably enhance retrieval efficiency without the loss of accuracy.