Evaluating Term-Expansion for Unsupervised Image Annotation

Abstract

Automatic image annotation (AIA) deals with the problem of automatically providing images with labels/keywords that describe their visual content. Unsupervised AIA methods are often preferred because they can annotate (virtually) any possible concept to images and do not require labeled data as their supervised counterparts. Unsupervised AIA methods use a reference collection of images with associated (unstructured, freeform) text to annotate images. Thus, this type of methods heavily rely on the quality of the text in the reference collection. With the goal of improving the annotation performance of unsupervised AIA methods, we propose in this paper a term expansion strategy that expands the text associated with images from the reference collection. The proposed method is based on term co-occurrence analysis. We evaluate the impact that the proposed expansion has in the annotation performance of a straight unsupervised AIA method using a benchmark for large scale image annotation. Two types of associated text are used and several image descriptors are considered. Experimental results show that, by using the proposed expansion, better annotation performance can be obtained, where the improvements depend on the type of associated text that is considered.