الفهرس 
AcknowledgementOnly 14 pages are availabe for public view
 |  | from | 108 |  |  |
| | | from | 108 | | |
Abstract
Low level features have been widely used in content-based image retrieval (CBIR) applications. Color and texture descriptors represent the majority of low level features implemented. A distance measure is then used to measure the similarity between images based on the descriptions of their low level features in order to find and arrange the relevant images retrieved. Pure dependence on visual features leads to what’s known by the semantic gap problem. Semantic understanding of images is frequently implemented in CBIR systems to overcome the problem.
In this thesis two different modes of features were extracted. The visual feature implemented was the Color and Edge Directivity Descriptor (CEDD) which combines in one histogram both color and edge descriptors. The other mode is the textual tags that were extracted from text surrounding web images. WordNet database is used as a lexical database to answer semantic requests about thesaurus and oncologic origins of words. A hybrid ranking algorithm is introduced to combine the visual features similarity and the tags similarity.
The Automatic Multimodal Content Based Web Image Retrieval System (AMIR) semantic CBIR system was implemented. About 24,000 of Real web images were collected in a database using 60 query terms and then they were tested for their visual relevance to these terms (ground truth images). As the database is large and expected to increase in volume; visual features were extracted and indexed using k-means clustering algorithm in order to speed up the retrieval process. About 300,000 Textual tags were extracted and weighted by the frequency of occurrence of both the words and their thesaurus. Finally the accuracy of the system retrieval was tested using comparison between the results of two different sets of experiments. The first used only the visual features and the second was using both visual and textual keywords.