A Bayesian approach to edge detection in noisy images

An adaptive method for edge detection in monochromatic unblurred noisy images is proposed. It is based on a linear stochastic signal model derived from a physical image description. The presence of an edge is modeled as a sharp local variation of the gray-level mean value. In any pixel, the statistical model parameters are estimated by means of a Bayesian procedure. Then an hypothesis test, based on the likelihood ratio statistics, is adopted to mark a pixel as an edge point. This technique exploits the estimated local signal characteristics and does not require any overall thresholding procedure.