The invention is a method for determining a segmentation into distinct regions of a set of coefficients representing at least part of an image in order to insert or extract a set of watermarking signals with a predetermined cardinal. The method of the invention generates a set of acceptable segmentations, calculates, for each acceptable segmentation, a probability of error on the detection of the watermarking signals inserted in said each acceptable segmentation, and selects at least a segmentation in the set of acceptable segmentations, on consideration of the probability of error subject to a constraint on the number of regions of the selected segmentation, with respect to said predetermined cardinal.