The present invention detects the presence of a watermark in-an image by using a multi-step process. First, the image is examined to determine which regions of the image have characteristics such that there is a high probability that a watermark signal can be detected in that region of the image. Next the regions that have a high probability that a watermark can be detected (in contrast to all regions of the image) are examined to find watermark data. In order to determine the probability of finding watermark data in a particular region of an image, the amount of "variance" in the intensity of the pixels in the region is first examined. For example a region that is entirely white or entirely black has zero variance in luminance. Such a region can not carry watermark data, hence regions with zero or low variance can be eliminated from further processing. Furthermore, if a high variance in a region is a result of the fact that the region has an abrupt border or edge between two highly contrasting regions, the high variance does not indicate a high probability that a watermark signal will be detected in the region. Therefore, after regions with high variance are located, these regions are next examined to look for regions with high edginess spread. Finally, regions with the high variance and high edginess spread are selected for further processing to detect watermark data. For those regions selected for further processing, the detection process can be enhanced by filtering the data with a two step process to increase the signal to noise ratio of the watermark signal. First a high pass filter (e.g. a Laplacian operator) is applied to each region. This filtering operation in effect established a new intensity value for each pixel in the region. Next a nonlinear operator (e.g. a signum function) is applied to the output from the first filter operation. The resulting data is examined to detect watermark data.