Automated analysis and prediction of accuracy and performance in ATR algorithms: II. Experimental results and system performance analysis

Automated target recognition (ATh) algorithm performance is often degraded by the propagation of sensor noise and input error through an algorithm's computational operations. In lieu of a detailed error analysis, algorithm perfonnance estimates are usually couched in terms of probability of hits (Ph) or rate of false alarms (Rfa). Although such measures may be useful to an algorithm's end users, they may not be meaningful mathematically to a designer or theoretician. That is, deficits in nonstandardized measures such as Ph and Rfa cannot usually be related analytically, in closed form, to noise and error at an algorithm's input. In Part 1 of this series of two papers [1], we discussed requirements for theory and software systems that support automated error analysis. We showed how the dataflow graph of an algorithm can be relabelled with error functions of each constituent operation. The resultant dataflow graph describes the error function of the algorithm. In this paper, we demonstrate the application of our novel error estimation software (called ERSIA -- ErrorReduction and Synthesis in Image Algorithms) to several practical image/signal processing (ISP) and ATh algorithms that were successfully employed in the detection of small targets in cluttered multispectral imagery [2]. For example, we analyze (a) the Adaptive Double-Gated Filter, (b) a statistical texture detector, and (c) gradient-based edge-enhancement algorithms. By perturbing the input with various types and degrees of noise (i.e., Gaussian, Lorentzian, and Poisson distributions), we show that ERSIA can accurately predict the output noise in each algorithm, and that these predictions agree with theory. We analyze the computational cost associated with each algorithm and with the automated error analysis, in terms of sequential (von Neumann) and SIMD-parallel computational paradigms. We also discuss the computational cost associated with the application of ERSIA to a given algorithm. Occasionally, this cost can be high, since ERSIA can be programmed to compute a two-dimensional error distribution for each image operand and scalar variable. Planned implementations of ERSIA on SIMD- and MIMD-parallel processors are discussed in terms of computational accuracy, cost, and portability.