Removing watermarks from stock photos is worryingly easy

Watermarking photos may be the go-to route to prevent copyright theft, but Google researchers have figured out a clever way to remove watermarks altogether, automatically. Watermarking, where a logo is placed somewhere on a photo to show who took it, varies in intrusiveness from a small caption in one corner to the sort of full-image obscuring lines, text, and other graphics that stock image companies often plaster their previews with. As the researchers found, though, the very ubiquity of that strategy might be its undoing.

Certainly, even a complex watermark can usually be removed by a keen Photoshop expert. It's a time-consuming process, however, which is where trying to figure out an automated alternative comes in. According to Tali Dekel and Michael Rubinstein's new paper "On The Effectiveness Of Visible Watermarks" which they recently presented at the 2017 Computer Vision and Pattern Recognition Conference (CVPR 2017), that's not quite as simple as you might expect.

In fact, they say, when faced with a single, watermarked image a computer might have just as much trouble as the human Photoshop pro does. "Visible watermarks are often designed to contain complex structures such as thin lines and shadows in order to make them harder to remove," they point out. "Indeed, given a single image, for a computer to detect automatically which visual structures belong to the watermark and which structures belong to the underlying image is extremely difficult."

Their solution takes into account the regularity of most watermarks. By giving a machine learning system access to a repository of different images that all use the same watermark – something which can be as simple as pointing it at one of the preview galleries at Shutterstock, Adobe Stock, or other similar sites which sell stock graphics – it can eventually identify the consistent watermark on each. As long as the watermark graphics are semi-transparent, even complex patterns including lines, color gradients, and shadows can be pulled out, and then readily removed by the system.

None of this is likely to come as welcome news to photographers or others, who rely on watermarking to help police copyright. Happily, the researchers have also made suggestions for more effective techniques that, while giving an end-result that looks the same to the human eye, is far harder for computer vision systems to identify.

That involves making the watermark itself slightly different each time it's used. By introducing very slight random warps, the same system they developed was unable to cleanly remove the watermark from the underlying image. It's a technique that works when other methods – like adjusting opacity or position on the frame – aren't sufficient for fooling the system.

Of course, if you want to put this into practice there's not much in the way of software to apply warped watermarks – at least, not yet. As the implications of this research spread, though, that seems something likely to change as photographers update their copyright toolkits.