Using Multiple Segmentations for Image Auto-Annotation

Abstract

Automatic image annotation techniques that try to identify the objects in images usually need the images to be segmented first, especially when specifically annotating image regions. The purpose of segmentation is to separate different objects in images from each other, so that objects can be processed as integral individuals. Therefore, annotation performance is highly influenced by the effectiveness of segmentation. Unfortunately, automatic segmentation is a difficult problem, and most of the current segmentation techniques do not guarantee good results. A multiple segmentations algorithm is proposed by Russell et al. to discover objects and their extent in images. In this paper, we explore the novel use of multiple segmentations in the context of image auto-annotation. It is incorporated into a region based image annotation technique proposed in previous work, namely the training image based feature space approach. Three different levels of segmentations were generated for a 5000 image collection. Experimental results show that image auto-annotation achieves better performance when using all three segmentation levels together than using any single one on its own.

More information

Identifiers

Catalogue record

Export record