Image processing consists of operations on an image whose result is another image. This article reviews many of these, placing particular emphasis on mathematical and statistical operations. Statistical image processing considers operations that can be evaluated according to their expected loss, which is a criterion that compares any estimated image to the true image. As the true image is unknown, a probability distribution (e.g., a Markov random field) is postulated for it, which is sometimes called the
prior distribution
of the image. Then, from the prior and the noise distribution of the observed image given the true image, the posterior distribution can be obtained from Bayes’ theorem. This distribution, of the true image given the noisy image, is high dimensional and may be difficult to calculate in practice. Statistically optimal algorithms that restore, segment, reconstruct, or extract features of an image are obtained by minimizing the posterior expected loss, where the loss function may depend on the operation being considered. Other topics covered include mathematical morphology, histogram equalization, boundary detection, feature extraction, and pattern recognition. Copyright © 2015 John Wiley & Sons, Ltd.