A color edge detector based on statistical rupture tests

This article presents a technique for detecting contours in color images based on abrupt change detections in parametric edge models. This abrupt change detection is performed on each line and column of the three component images (red, green and blue images). This technique consists in computing at each pixel a change point criterion based on a statistical change point detection. The criterion value is compared to a threshold. Two statistics models are used: the generalized likelihood ratio and the divergence statistic. The performances of these two models are about the same when applying decreasing exponential weights to the data. When there is a abrupt change on a pixel the gradient value is obtained by making the difference between the two grey-level averages on the two sides of the pixel. Finally, Di Zenzo's (1986) combination is performed in order to get the color gradient.