An Evaluation of the Effectiveness of Image Features for Image Retrieval

Retrieval effectiveness in image databases depends significantly on the features and the distance model utilized to evaluate the similarity of the images. Features must be extracted from images and stored in the database. Since the features should be stored at the time of data entry, it is extremely important to determine which features ensure the best retrieval performances. In this paper a comparison of the most widespread automatic indexing techniques and their performances is presented. The image reference set, necessary for performance comparison, is obtained by including in the database frames extracted from video shots. Frames extracted from a shot are different but have the same semantic content. One of these frames is utilized as an example in a query; the indexing effectiveness is assessed from the frames retrieved. The most relevant features prove to be the angular spectrum, the Hough transform, and the color histogram, followed by local features such as local luminance pattern directionality.