Evaluation of FLIR sensors for automatic target recognition using an information-theoretic approach

Abstract

This paper describes the development of an information-theoretic image measure for Sensor evaluation under the contract to the United States Air Force. While current approaches are based on human perception models, a need exists for evaluation of sensors for ATC/ATR systems. Such an evaluation should be performed in terms of the probabilities of detection/identification and false alarms independent of the idiosyncrasies of the specific ATC/ATR algorithms. Such an approach based on the informationtheoretic content of images for the target vs. background separability is being developed and applied to evaluating sensors using the Tower Test data collected at the Wright Laboratories. INFORMATION THEORETIC VS. EMPIRICAL APPROACHES TO ATR SYSTEMS EVALUATIONAn ability to predict the performance of ATC/ATR algorithms and sensors is critically important for a number of applications. Tactical Decision Aid (TDA) systems need this capability for optimal settings of ATC/ATR parameters. Mission Utility Models need this capability for predicting mission success. Data collection activities need this capability for optimal experimental design. The performance of each specific ATC/ATR system depends on multiple factors such as the overall structure and concept, sensor, the performance of algorithm subcomponents, quality, adequacy of the training data set, and the accuracy of the tuning algorithm that sets parameters for specific missions. One way to model the performance of an ATC/ATR system is to obtain empirical data on the performance by using a large data set of imagery, and then, fit a (nonlinear) regression model to these performance data as a function of IM. On the other hand, it is well recognized that in each image there is specific information pertinent to the separation of target from background and to target identification, and the amount of information determines a bound on the best possible performance of any algorithm.Various applications have their own performance evaluation needs. While TDA and mission utility models should be modeling available tactical algorithms, data collection and sensor evaluation should not depend on idiosyncrasies of specific algorithms, and should be based on information-theoretic considerations. Similarly, algorithms' performances should be evaluated against information-theoretic performance bounds. In this paper we develop the Information-Theoretic Image Measure (IT IM) which utilizes the information content of an image for the purpose of sensor evaluation. Current approaches to sensor evaluation art' based on models of human operators' ability to detect and identify targets. Their results are usually expressed in terms of maximum detection/recognition range. Development o f sensors for Automatic Target Cueing and Recognition Systems (ATC/ATR) s h a1 1 be performed in terms of probabilities of detection and false alarm and shall satisfy needs of existing and future ATC/ATR systems. Characterization of sensors in terms of probabilities of detection and false alarm can be obtained by eval...

Aided and automatic target recognition based upon sensory inputs from image forming systems

“…Attempts to create additional image metrics based on information theoretic approaches by Clark et al [10], [11] have not met with success. The failure in our ability to expand the available metrics is probably due to the enormous variability of the input scene [12].…”

“…Although many pixels may be inconsequential and many combinations physically unrealizable, a huge library of possible candidate decisions must be considered by an ATR processor. In a dynamic scene where the image may change at TV frame rates (30 frames per second), there are potentially 10 8 digital values per second that must be processed. Obviously, ATR requires an enormous amount of processing power, speed, and memory and necessitates smart software strategies.…”