Image Complexity Metrics for Automatic Target Recognizers

Designers of automatic target recognizers (ATR) need measures of image complexity to compare the performance of different ATRs. An image complexity metric should provide an a priori estimate of the difficulty of locating a true target in an image. An ideal image metric is a mapping from the set of all images to a finite real interval. The extrema of the interval indicate extrema in difficulty. The mapping must be monotonic in probability. An ideal image complexity metric is independent of specific ATRs and targets. This is, of course, an impossible ideal. In the context of ATR design, complexity must be linked to the difficulty of the task. But, tasks that are difficult for one ATR may be easy for another and vice-versa. Therefore, there can be no completely ATR-independent complexity measure. It is possible, however, to define a class of ATRs based on the similarity of the image features they use for detection. Within such a class, it could be possible to define a nearly ideal metric, since the image characteristics which frustrate one ATR present similar difficulties to the others. This metric would be a measure of image features. But, its computational definition would derive from the definitive attributes of the ATR class. In this paper, we review recent ideas about complexity in general, and we review some measures of image complexity in particular. We demonstrate that a significant number of ATR algorithms in the public domain literature share similar computational attributes. We use these attributes to define an ATR class. We then propose an image complexity metric for the class.