Refining Image Annotation Based on Object-Based Semantic Concept Capturing and WordNet Ontology

This paper presents a novel approach to automatically refining the original annotations of images. An existing image annotation method is used to obtain the candidate annotations for an image in advance. Then, low-level features are extracted automatically from all blocks in the image to construct a suitable multi-feature space. Next, the image is divided into nonoverlapping block-based structures and a block-based structure clustering algorithm to capture the semantic concept of object as accepted annotations is proposed. Based on these accepted annotations, the irrelevant annotations are pruned according to the semantic similarity in WordNet. Experimental results on the typical Corel dataset show that the approach outperforms the existing image annotation refining techniques.