Avoiding Class-conditional Independence Assumptions in Image Classification

Much published work on contextual image classification is based on an assumption of class-conditional independence (CCI) of the measurement data equivalent to assuming that an ideal classifier will recover the underlying true scene, with errors evenly and randomly distributed as white noise. This paper proposes a simple alternative model which, it is argued, is more realistic in many applications and upon which useful theory can still be built. The new model is then used to investigate the effect on the accuracy of an object classifier which makes the CCI assumption in a domain where it is not valid.