Research On Web Image Retrieval Based On Web Information And Image Feature

The number of website is steeply rising due to the rapid development of Internet and information technology. The age of mass information resources has entered due to widely using new media, such as Twitter,We Chat and micro-blog. In these resources, not only including the text data are easy to processing, also including a large amount of multimedia data such as images, videos. Compared with the text, the image more intuitive, vivid, coupled with the popularity of portable image acquisition devices such as smart mobile phone, the data quantity of digital image appeared high speed growth, almost doubling each year. It is an urgent task to how to quickly and accurately retrieve Web images from massive cyber source.Web image has the characteristics of high information, huge quantity, unstructured, diverse high-level semantic and nondeterminacy. So the relevant achievements of the field of information retrieval are hard to be applied to the Web image retrieval. It has the important theory significance and the practical value to study a high performance Web image retrieval system using theories and technologies such as digital image processing, pattern recognition, machine learning and information retrieval,and to provide stable and reliable and high accurate image retrieval service for users.Due to web image is embedded in the webpage,sothere are two types of information in Web image: one is visual information of web image, two is the text information of webpage which contains web image.Research on Web image retrieval is mainly based on the two types of information. It includes three aspects: TBIR(Text-based Image Retrieval), CBIR(Content-based Image Retrieval) and ABIR(Association-Based Image Retrieval). TBIR extracts keywords from the text of webpage which contains web images, and the keywords are used for image indexing and retrieval. CBIR extract keywords from the low-level visual information of web image. Information of text and image is used for image indexing and retrieval in ABIR. Obviously, ABIR has more superiority due to the use both of text and image. However, how to combine the two types of information is still very difficult because of the different structure between them(One is the high-level semantic keywords, and the other is the low-level visual information). A sample or sketch is required for directly image retrieval using low-level visual information, which is great inconvenience for user. "Semantic gap" is another problem in translating low-level visual features into high-level semantic keywords. The study progress of ABIR is slow due to the two problems. The further study is necessary.To address the above issues in ABIR, a Web image retrieval model based on webpage information and image features is proposed in this paper. The information source of Web image retrieval is extended from two types(Text information and image features) to three types(Text information, image features and text in Web image). The image visual features is mapped to the high-level semantic keywords by automatic annotation, therefore, "semantic gap" problem is solved. Keywords from three information sources are fused together by Bayesian network, and the problem fusing different structure is solved.Term similarity computation is the base to solve "Semantic gap" problem, and is also the premise to fuse the three information sources in proposed model. Therefore, the method of term similarity computation based on How Net is detailed in 4th section.The text in Web image includes two categories: scene text and artificial text. The two categories are great interrelated with the content of Web image and the high-level semantic keywords. It has important significance that the text in image is usedas one of the image retrieval information source. The texts in image arerecognized by the proposed text recognition algorithm based on Stroke Width Transform.Automatic image annotation is an effective means to solve the problem of "semantic gap". The automatic annotation proposed in this paper includes the following steps: firstly, candidate keywords set are generated by three classical automatic annotation models. Secondly, the cohesion among candidate keywords and the coherence between the keywords of webpage text and candidate keywords are computed by term similarity computation. Thirdly, the keywords whose value of cohesion and coherence are low are filtered out from the candidate keyword set. Meanwhile, the cohesion and coherence are usedas the weight of keywords, which is prepared for information fusion based on Bayesian network.Bayesian network has an inherent fusion capability of multiple information sources, and it is a kind of inference network based on probability and uncertainty. The key using the inference network is how to determine the initial probability and conditional probability. In this paper, the initial probabilities are determined by Page Rank, and the conditional probabilities are computed by TF/IDF. Thus all the parameters of a Bayesian network are determined. Experimental results demonstrate that the Web image retrieval method proposed in this paper has higher image retrieval quality.