Research On Image Retrieval Based On Text And Visual Information

Digital images are produced and published at a really impressive speed. Images play a vitalrole in our daily life for the reason that information contained in them tend to be much more thanthat in the text. Currently, it becomes a hot topic in the field of image retrieval that how tomanage the large image data base more effectively and return the most satisfied images back tousers. There are mainly two kinds of models for image retrieval: Text-based image retrievalsystem and Content-based image retrieval system. The former realizes retrieval by using textinformation around the image, there are two main drawbacks which restrict the development oftext-based image retrieval: Keywords annotated to an image are various due to the diversity ofpersonal individuals, as well as, the cost caused by manual annotation is unbearable. A lot ofcombined models are proposed by Researchers to highlighting the favorable aspects andavoiding the unfavorable aspects of the two methods, and the retrieval performance hassignificantly improved. This paper focused on image retrieval and studied the followingproblems: How to substitute manual operation by annotate images automatically, how to producemore representative visual features and how to combine the text and visual information togetherin the retrieval progress.（1） The existing automatic image annotation technology often divided an image intoseveral independent blocks, then, the keywords are distributed to each block according to thesimilarity between them. But the types of images are various in our daily life, some images(scene image) contain quiet simple content in comparison with others. Segmentation will notonly result in excessive workload but also noisy annotations. Thus, in this paper, we use differentfeature extraction method based on the complexity of the image, at the same time, bayesianclassifier also be combined together to annotated an unlabeled image automatically.（2）Traditional content-based image retrieval system usually concerned the visualinformation only while ignored the information lies in the surrounding text. In this article atwo-stage retrieval model is proposed to tackle the problem. Firstly, images are ranked based onthe semantic similarity lies in the surrounding text and the irrelevant images are filtered out. Onthe other hand, the more relevant images are reserved to form the reduced image data base.Secondly, contend-based image retrieval performed on the reduced database, and the retrievalreturned to users as the final results. In this paper, canonical correlation analysis algorithm isapplied to learn the relationships between different visual features, and Word Net and IC modelare performed to analyze the semantic information which lies in the surrounding text.（3） Traditional image automatic annotation system usually annotate an unmarked imagewith keywords attached to images which are most relevant to it, and ignored the high levelinformation lies in the text. In this paper an image annotation refinement method is proposed.KNN algorithm is performed to select the nearest neighbors of the unlabeled image and allocatedifferent weights to different neighbor according to the similarity between them. WordNet andrelated technology are applied to generate the base-label and the expand-label collections to eachof the unmarked image. The proposed method considered not only the content similar keywords but also the semantic similar keywords. Experimental results show that the performance ofannotating system proposed in this paper is improved.