Comparison of Classification Algorithms on MSTAR Data Using Risk-Based Empirical Statistics

摘要

As the sum of the products of cost and probability for all types of classification decisions, total classification risk for a classification system is easily calculated. Empirical risk data produced by Monte Carlo simulation of the battlespace lends itself to statistical description of total classification risk for comparison with other classification systems. Families of classification systems are created using Probabilistic Neural Nets (PNN) acting on the Moving and Stationary Target Acquisition and Recognition (MSTAR) mixed targets data set. The spread parameter of the PNNs serves as one threshold distinguishing the PNN classification systems from one another, and a second parameter is a cropping proportion used in processing the image data. Using computer simulation, a warfighter can choose a threshold that minimizes risk under the assumption of temporarily fixed costs.