Research On Web Image Retrieval And Classification

With the development of computer and network techniques as well as widely used multimedia, images on the WEB increasing at an alarming rate. The use of image information makes its way into every social fields and images has become very important public digital information. More and more people can get this kind of information conveniently. At the same time ,lacking of images is no longer a problem .Instead , the issue becomes how to location necessary information among such a massive image database .So research on how to organize ,manage and retrieval the large WEB image database has great value to future Internet service .In this paper, the necessity of building image retrieval system on scientific document database is analyzed based on research on status and applications of WEB image retrieval .In the part, a method which based on image segment is proposed to extract images from PDF documents .This method is proved to be effective and popularity and has a high precision .Next, titles, abstracts, keywords and surrounding text are extracted. Different combinations of them are used to index the images. Finally, a text-based image retrieval prototype system on scientific documents database is built .The system is very complimentary to current retrieval system on scientific documents database.In addition, a rule based method is proposed to extract metadata for WEB images after looking into the general review and key techniques in WEB image retrieval and classification area .The method can be employed to solve the issue how to make use of text information in HTML pages in the process of WEB image retrieval and classification .It is tested on a HTML page database which includes many pages from several portals .The experiments indicate that the metadata generated by this method is very descriptive. Rest on the previous work, an algorithm integrated MPEG-7 color descriptors and metadata to classify WEB images is raised. In this algorithm, image content feature and text feature are quantified together. The experiments show that the image content feature and its text feature are complimentary with each other and combining them has a great impact on WEB image classification. Simultaneously, it's proved that the rule-based WEB image metadata generation method is effective.