Multi-Feature Based Web Image Annotation System

In recent years, digital image has become more and more popular along with the improving technology for digital media and the convenience of internet. Many digital images become usable every single day. It does make sense to design an image searching system that could satisfy users' image searching requirements. Now there are two techniques of image searching. The first is based on the keywords known as TBIR(Text-Based Image Retrieval). The second is based on content of the image known as CBIR(Content-Based Image Retrieval). The difference between two techniques is the annotation method for images.The image annotation method of TBIR has two different ways: one way is to annotate the image,by manually selecting keywords; the other is to generate keywords for images using automatic annotation system.The content-based image searching is to obtain visual features from image source directly, such as color, texture, shape to label the content of image. When we search an image, one or more images of which the visual features are similar to the image searched will be returned as the searching result. It involves a lot of computation. Moreover, when these visual features are finally reflected to semantics, the annotation results cannot satisfy users' understanding because of images' complication and uncertainty of the semantics of the annotation words.This paper has brought forward a new image annotation method to overcome the disadvantages of the existing ways above and improve the quality of image searching on internet. The new image annotation method utilizes the technology of NLP, text classification and self-adaptive web text extraction and then annotates the web images based on the multi image correlative features. The correlative features includes: the URL of the image, the image context, the theme of the web page and etc.The emphasis of the work is to design and implement a multi-feature based web image annotation prototype system. This includes two main parts. One is extracting web page features part in which two algorithms are designed: DOM-tree-based image context extraction method and vision-and-rule-based web content extraction method. The other is generating image annotation part in which using the correlative features in the web and text classification technique to generate annotation correctively.