الفهرس Only 14 pages are availabe for public view
 |  | from | 89 |  |  |
| | | from | 89 | | |
Abstract
Web image classification is one of the most important and hot topic research areas. The web has a huge volume of unstructured and distributed documents full with multimedia data such as image, voice and video in addition to the text that are related to a variety of domains. In addition, the web is full of noise that will certainly harm the classification process; so considering one base for the classification task will be extremely difficult. The demand of categorizing and indexing this huge repository of information is a challenging task.
Most commercial search engines and web directories (also called portals or catalogues) rely on text only in their categorization. process where the user makes a key-word based query that usually results in a long list of retrieved documents with a low precision in multimedia results. This is due to the dependency on text-based not content-based image in the categorization process and also because of the huge size of the web.
We believe that building a domain oriented classifier or vertical classifier which is concerned with a specific domain (like Military in this research) will certainly minimize the scaling problem. Moreover, we will be able to build catalogs based on image content and other complex data types. This system may be beneficent in a Variety of applications which depend on finding similar images.
Instead of using a flat classifier to combine text and image classification, we divide the classification task into smaller classification problems corresponding to the branches in the classification hierarchy, we apply a text-based approach and an image based approach on different levels separately. Each of these subtasks is thus significantly simpler and their features are better discriminated than treating them as a unified task in a flattened classifier. Thus the resulting models are more efficient and achieve a higher accuracy.
In this research we propose a technique to build a hierarchal image directory based on the color signature image feature and the association between image content and non image content. The Earth Mover’s Distance (EMD) is used as a method to evaluate dissimilarity between two multi dimensional distributions in some feature space; it is used as a similarity measure between two image signatures in the K-nearest neighbor algorithm to find the similarity between images.