Performance Bounds for ATR based on Compressed Data

by Avinash Jain, Pierre Moulin, Kannan Ramchandran and Michael I. Miller

This work conducted within the Army's Center for Imaging Science (CIS)
focuses on quantifying performance loss when Automatic Target
Recognition (ATR) systems operate on compressed image data. We
consider the problem of target detection based on sensor data
compressed using transform-based coders. Information-theoretic
distances such as Kullback-Leibler and Chernoff distances are used to
bound detection performance. Detection problems under known and
unknown orientation of target are considered. Analytical expressions
are derived for a simple sensor noise model. This study provides a
systematic framework in which to study degradation in detection
performance due to lossy compression. Monte-Carlo simulations for
detection of a T62 tank in additive white Gaussian noise are presented,
and actual detection performance is compared with Chernoff bounds.