Abstract
This paper proposes a keyword hierarchy construction method and its application to automatic image annotation. This method takes advantage of the characteristic that keywords with high-level semantic meaning tend to high visial diversity across the corresponding image set. Based on this characteristic, the keyword hierarchy is constructed by using the following approach: First, a visual feature clustering method is applied to a set of images that are annotated with a target keyword. Then, a novel criterion that represents visual diversity of the keyword is introduced. Specifically, this criterion is calculated using both intra-cluster and inter-cluster visual similarities. Finally, the keyword hierarchy can be constructed by sorting the keyword criterions. Based on the obtained keyword hierarchy, low-level keywords are first estimated from visual features, and then high-level keywords are provided by using semantic relationships of the low-level keywords. This will improve performance of image annotation.
