Key Technologies Of The Internet Image Search Engine Based On Multi-cues Integration Study

As the rapid development of network bandwidth and CPU speed, the embedded images in HTML pages prevail nowadays. For its amazing function in describing the content, attracting the attention and even entertaining, images become substantially important in web pages. These images form a very rich gallery for users. What's more, the semantic meanings of these images are well presented by the surrounding text and links. However, the image search engines do not dig up them in a proper way, which leads to low recall and precision. In addition, the nature of language brings in ambiguity in just using keywords to represent the image. There comes to many categories of images match the keyword. Unfortunately, the image engines don't give a classification for these totally different images, which makes it very annoying for users. Aiming in the usage of Digital Library, this thesis focuses on the improvement for image search engines, mainly basing on two techniques, multicue WWW image annotation and WWW image classification. On the basis of these, we have implemented a WWW image search system: ISearch. All the above work is beneficial to Digital Library and the related field of information retrieval. We first introduce the background of our work, the weaknesses of the traditional technique and the construction of this thesis. Then we trace back the history of information retrieval and some typical image search systems. After that, we introduce four WWW image annotation methods, based on low level features, web page tags, overall word frequency and local word frequency. Then we put forward our method of image annotation. In Chapter 4, we do research in image classification, including a couple of similarity models, vector scaling, clustering and ranking. Chapter 5 mainly introduces our image search system ISearch, which bases on the techniques of Chapter 3 and Chapter 4. We sketch its construction and implementation, then following its performance. Finally, we give a conclusion and talk about the future of image retrieval.