Developing texture-based image clutter measures for object detection

Abstract

Automatic object detection is one of the basic tasks performed by an image understanding system. Object detection approaches need to perform accurately and robustly over a wide range of scenes. Although a number of detection approaches have been developed and reported, a need remains for standards by which to judge the relative merits of such approaches. Image characteristics, object characteristics, and detection methodology are recognized as the main variables affecting object detection. A basis for their quantitative analysis and evaluation is developed. This research keeps object detection methodology constant while varying image and object characteristics to develop a set of quantiative standards. This requires an ability to derive a quantitative measure for the observed in an image. A performance index for object detection approaches, as a function of scene nature, is valuable. Current approaches to image clutter or quality characterization are studied and a new measure based on image texture content and object characteristics is proposed. An extensive set of experimental studies is utilized to evaluate this texture-based image clutter (TIC) measure. TIC is shown to be better suited than other reported clutter measures because of its ability to accurately quantify perceptual effects and to serve as a robust indicator ofthe object detection and false alarm rates as a function of image clutter.