Unsupervised smooth contour detection

Abstract

An unsupervised method for detecting smooth contours in digital images is proposed. Following the a contrario approach, the starting point is dening the conditions where contours should not be detected: soft gradient regions contaminated by noise. To achieve this, low frequencies are removed from the input image. Then, contours are validated as the frontiers separating two adjacent regions, one with signicantly larger values than the other. Signicance is evalu-ted using the Mann-Whitney U test to determine whether the samples were drawn from the same distribution or not. This test makes no assumption on the distributions. The resulting algorithm is similar to the classic Marr-Hildreth edge detector, with the addition of the statistical validation step. Combined with heuristics based on the Canny and Devernay methods, an efficient algorithm is derived producing sub-pixel contours.