Abstract—Prediction of separability of potential targets classes is an important issue in ATR systems. Unfortunately despite ATR systems have been adopted since the 1960's, the definition of class-separability is still an open question. Kolmogorov Appropriate Prediction of Separability (KAPS) seems to be a powerful tool to predict the degree of separability of classes. The article is addressed to understand how reliable the metric is in the degree of separability assessment. The methods used for evaluating the performance of metric are order statistics analysis and Monte Carlo simulations. The analysis consisted in understanding how many samples are necessary in order to get a reliable value of KAPS metric. As demonstrated by results, the KAPS value depends on the number of samples, therefore it is suggested to use a number of samples of the order 103at least.

Index Terms—Automatic target recognition, machine learning, KAPS, class-separability, degree of separability.

The author is with European Commission, Joint Research Centre, Institute for the Protection and Security of the Citizen, Maritime affairs, Via Enrico Fermi 2749, I-21027, Ispra (Va), Italy.

Cite:Giovanni Marino, "Study on Reliability of KAPS Metric as Degree of Separability for Classification Problems in ATR Systems," International Journal of Computer and Communication Engineering vol. 5, no. 1, pp. 23-30, 2016.