An investigation into noise-bound shadow detection and removal

Abstract

Noise is an unavoidable contaminant in any non-trivial image. It is usually identified as a limiting factor in the performance of shadow-removal algorithms, but little is done to reduce its negative impact. The typical method to counter noise effects is to employ arbitrary or empirical thresholds somewhere inside the algorithm, with values chosen to maximize the shadow-removal performance. However these thresholds can be objectively calculated from the noise statistics for a particular pixel value. We present a method of shadow-removal whose internal parameters are adaptively set by noise statistics such that the algorithm is free of any empirically set threshold. Experiments indicate that the performance of the new method is approximately equivalent to that with an empirically-fixed threshold, though an area of improvement has been identified that could significantly boost the accuracy of the new method.