Research On Domain Ontology-based Semantic Image Retrieval

The progress of computer technology and internet technology results in theexplosive growth of multi-media data such as images. How to effectively and quicklyretrieve required images from large-scale image database is a hot research. Recent years,semantic-based image retrieval techniques have been concerned in order to attempt toextract knowledge from these images, enhancing the retrieval performance. Asemantic-based framework is presented to promote semi-automatic annotation andsemantic image retrieval using multimodal cues such as visual feature and text captions.In addition, a hierarchical structure of the semantic model allows data to be shared andsupports multi-semantics (polysemy) for concepts. In this thesis, we proposed domainbased semantic image retrieval model, the model realizes the transformation ofunstructured vector space model to a hierarchical ontology model using a technique.Under the ontology framework, image retrieval is completed by concept matchingrather than simple low level feature similarity, system’s retrieval performance (higherprecision and recall) is improved using this retrieval mechanism. To this end, the mainwork of this thesis as follows:①Domain ontology construction. We acquire expert knowledge related tosports(such as field event) from Olympic website, extracting relevant concepts, relationsand their property.②Bag of visual word Generation. SIFT algorithm will be used to detect keypointsand perform vector quantization, and then using a novel clustering algorithm--semanticlocal adaptive clustering(SLAC), clustering the detected keypoints. Every generatedcluster will be a visual word, all the visual words construct a bag of visual word.Because SLAC takes keypoints’ space location and term weighting into account, so thequality (strong representation for visual content) of generated visual words areimproved.③Redundant visual word detection and removal. We need to detect and removethose that are useless to represent the content of visual data and can degrade thecategorization capability, in order to reduce the computation expenditure and improvethe retrieval performance of system.④The best transformation of unstructured visual word model to hierarchicalontology model is researched, in order to resolve the synonymy and polysemy of visual word.⑤Some uncertainty of visual content interpretation caused by several kinds offactors. We attempt to resolve these uncertainties to enhance the categorization andretrieval performance of system.We research these problems to construct the image retrieval model. The lastexperiment results shows that our method not only effectively represent the visualcontent of image, but also significantly improve the retrieval performance of system.