Methods for Detecting Watermarks
If a watermark is visible in spatial domain, and it is not scaled or rotated, then given the watermark, one can use cross-correlation to locate it. The point with the highest magnitude of cross-correlation is most likely the location of the watermark.
Matched Filtering
Original Image
Watermark
Original Image and the watermarked image side by side. Notice the watermark in the top left corner of the right image.
The peak magnitude of cross-correlation is found to be 0.9818. After taking account of the size of watermark and their respective offsets, a box can be drawn around the watermark found:
This method immediately breaks down when the watermark is rotated. The program tries to match the most evenly-lit area in the picture, possibly due to its similar color tone with the watermark. The peak magnitude of correlation is merely 0.4811, an unsurprising failure. Matched filtering can only deal with shifts, not rotations.
Frequency Domain Processing
One way to potentially be able to recover the watermark using frequency domain processing is by using edge detection. Using the first example, the edge detection method returns the following image (on the bottom right of this plot):
Clearly, applying the simple edge detection formula shows the watermark that was used on the original image. There are a couple of ways to render edge detection of watermarks meaningless. The best way (in our opinion) is to apply the watermark as an invisible watermark, which, for obvious reasons, would make it extremely hard for edge detection to identify the watermark. Another way to make the watermarked image resistant to edge detection is to choose a watermark such that it is an altered version of the original image (such as a grayscaled version), which would mean that edge detection would not be able to distinguish between the original image and the altered version. An example of the football image used throughout this report and a watermark which is a grayscaled version of it is shown below:
Lastly, the steganographic image (the one where the watermark is a ‘hidden message’, or a bunch of the letter ‘T’ in this case) is obviously resistant to edge detection, as no edges are being changed from the original image.