الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 136 |  |  |
| | | from | 136 | | |
Abstract
With the popularity of digital cameras, and the rapid developments of related Internet technologies, especially social services such as Flickr and Facebook, a huge amount of images are produced and distributed every day. As a result, the requirements for images retrieval systems are growing rapidly.
There are two common approaches for images retrieval systems; Text Based Image Retrieval (TBIR) and Content Based Image Retrieval (CBIR). The limitation of TBIR is the subjectivity of human perception and lack of semantic, which the limitation of CBIR is the semantic gap between image low level and high level of semantic cotenant. The aim of this work is tackling the mentioned problems of TBIR and CBIR. This thesis proposes an improved automatic image annotation and retrieval system based on semantic content.For the annotation process, the proposed system aims to automatically annotate unlabeled image and to filter out all irrelevant (noisy) keywords from user’s image annotation. To reduce noisy keywords, a semantic ranking concept score is computed for each candidate keyword associated with the test image. Keywords with very low scores can be removed from the final keyword set.After that; for each image, all its description keywords are semantically analyzed and frequently mined. Given an image, the proposed systems firstly initialize the candidate keywords set from its similar images. The candidate keywords set is then expanded by considering the implicit multi-tag associations mining.The proposed system provided three methods for image retrieval; firstly, the user select a region of query image and all images containing this region will be retrieved. Secondly the user enter the text query to the system and the will retrieve all image which is semantically related to the query tags, while the last method retrieve all images to the user which are similar in content and low level features. The experiments show that the proposed system outperforms TBIR and CBIR systems in annotation and retrieval performance.