Google’s latest feat could be seen as morally ambiguous at first. See, the team managed to create a software that removes watermarks from pictures flawlessly. Why would they develop such a counter-intuitive thing for professional photographers and stock photo agencies? Only to open their eyes and make them take precautions against it #softwaremagic
Google’s researchers published a paper called “On the Effectiveness of Visible Watermarks” where they showed how a computer algorithm can easily bypass common watermark strategies today. And if that can be done by an algorithm already, millions of pictures are at risk of being used without credit.
To do so, they trained the software with stock images until it learned the patterns of various popular watermark. Then, by separating the watermark elements from the rest of the picture (multi-image matting), the software could recover the original.
All this took place without human intervention or other information about the watermark. The end result looked flawless because the algorithm took into account the structure, opacity, structure, shadow or color gradient effects of the watermark.
So, what can be done to trick the computer? The researchers’ advice is to add “random geometric perturbations to the watermark”. Basically, if the text or logo used is warped, the computer has a harder time estimating the warp field and will usually remove just parts of the watermark, but not enough to make it reusable.