| In this thesis, research is performed on intelligent image content analysis to achieve efficient image database management. We investigate the use of image content analysis and interactive learning techniques as a tool of understanding image content, organizing image databases and choosing appropriate low level features as well as semantic meanings for image indexing and retrieval.; Image content analysis and semantic prediction techniques are utilized to achieve the coarse and fine level classification, respectively. In coarse classification, images are categorized by distinctive contents and indexed by the most suitable features. Image semantics are further predicated and images are finely classified by semantic meanings. Later, a scheme that learns image similarity and categories from relevance feedback is presented. First, we categorize each image by predicting its semantic meanings. During the retrieval process, users are allowed to confirm semantic classification of the query example and evaluate retrieval results by relevance feedback. By analyzing the feedback information, the system learns both image similarities and semantic meanings. In image similarity learning, the retrieving results are refined by modifying the similarity metric. At first, we don't have sufficient input information from users so that only weighting factors of feature vectors are justified. With more information available, we use the decision tree to learn relevance feedback and then apply the derived new similarity metric to update the retrieving process. The above refinement is conducted for a particular example containing a reasonable size of images. After recording feedback information of examples from different categories, we automatically obtain handy training data of image categorization. Image semantic learning is performed by using the decision tree training algorithm. |
