A note on evaluation of image recognition systems

Abstract

The ability to systematically evaluate the results of automated image processing systems has been problematic. It remains so. In this work we address two issues in testing: 1) the use of incomplete and inaccurate reference data; and 2) we introduce a new, somewhat faster method of evaluating systems that recognize linear structures. We illustrate the latter using road and lines-of-communication recognition systems. We describe past work on evaluation systems and compare their strengths and weaknesses relative to the current work. © Springer-Verlag 2003.