Research Of Region-based Automatic Image Semantic Annotation Algorithm

Today a large amount of information appears in the form of images. Consequently, ininformation management there comes the demand of image retrieval. A image retrievalsystem should be able to effectively find in the image database the images that the userrequires. Retrieval based on semantics is a natural way to do this. The user provides someindex words and the system finds all the images whose contents are related with thesemantic meaning of these words. This method requires that all the images in the databaseare annotated with keywords. Manual annotation requires too much efforts and is prone tosubjective bias. Therefore automatic semantic annotation for images become a hotresearch topic.The thesis first introduces the development of image retrieval systems and thedevelopment of automatic image annotation algorithms. The region-based automaticimage annotation algorithm is proposed. The work includes:The image segmentation method is studied. The clustering based segmentationalgorithm and uniform segmentation method are compared. The results show that theuniform segmentation method is computationally simple, easy for feature extraction andgives better annotation performance. Moreover, for uniform segmentation, a largernumber of regions give better annotation performance.The feature extraction methods are studied. Proper color and texture features areused to represent the image.The Gaussian Mixture Model (GMM) is used to represent the distribution of featuresof each semantic class. The parameters of GMM are learned by the iterative EM algorithm.Then the posterior probabilities are calculated indicating the possible semantic words foreach text image.A new decision method that considers the semantic relevance of each region is proposed to get the final annotation result. Compared with traditional method that onlyuses the posterior probabilities, the proposed method has higher recall, precision and Fvalues.