Initially this work focuses on identifying various sizes of semantic image feature buildingudblocks which can be used to represent an image as a bag of semantic image features. Then audmodified Bag-of-Visual Words (BoW) representation method is introduced which is differentudfrom the typical histogram-based approach. An image is converted to a document and then to audbinary signature to have a control over retrieval speed by reducing the feature space. Furthermoreudconverting an image into a document and to study how well it behaves with text processingudtechniques which simplify the CBIR has never been atempted. Therefore, this research forcusesudon finding the best sub-image size which can be used in BoW. Sub-image is the main buildingudblock of this research and hence it is characterized by color, texture and shape features usingudBoW with the feature index and the index of the nearest cluster center. Then Random Indexingud(RI) is introduced to CBIR by applying RI on the generated text file. The performance of theudproposed approach is evaluated using three benchmark datasets for quality, speed and robustnessudwhich confirms that the proposed approach has a high potential to retrieve correct images whichudin turn can be extended for a large collection. System performance is compared with existingudsystems in the literature and the results prove that our approach has superior performance overudthe other systems.
展开▼
