Feature Selection to Relate Words and Images

ABSTRACT

Image annotation, i.e. mapping words into images, is currently a major research problem in image retrieval. In particular, images are usually segmented into a number of regions, and then low-level image features are extracted from the segmented regions for annotation. As the extracted image features may contain some noisy features, which could degrade the recognition performance when the number of keywords assigned to images is very large, image feature selection needs to be considered. In this paper, a Pixel Density filter (PDfilter) and Information Gain (IG) are used as the feature selection techniques. By using Corel as the dataset, 10, 50, 100, 150 and 190 keywords annotation are setup for comparisons. The experimental result shows that PDfilter and IG can increase the precision of image annotation by colour or texture features. However, they do not enhance the annotation performance by the combined colour and texture features..