http://www.digitaljournal.com/tech-and-science/technology/google-creates-ai-algorithm-that-removes-watermarks-from-photos/article/500335

Google creates AI algorithm that removes watermarks from photos

Posted Aug 18, 2017 by James Walker
Google has created an AI algorithm that's capable of removing watermarks from images with almost flawless accuracy. The discovery could allow cybercriminals to steal stock photographs and professional images. Google provided advice to combat the issue.
Google entrance
Google entrance
Google
Research scientists Tali Dekel and Michael Rubinstein detailed their method in a blog post this week. The technique exploits the predictability of watermark styles and positions to allow an AI to automatically detect them. Watermarks, prolific on stock photo sites and the galleries of professional photographers, are meant to prevent unlicensed images being used in third-party works.
Google s AI can remove watermarks from images
Google's AI can remove watermarks from images
Google
READ NEXT: Google's G Suite gets new collaboration features in latest update
The trouble is that the watermark is applied uniformly to each photo. Regardless of the website, the watermark is almost guaranteed to be a transparent strip of text centred over the image. Adobe Stock, Shutterstock, Thinkstock and other major providers all have watermarks that remain unchanged between individual images.
This makes it relatively trivial for modern artificial intelligence to detect and remove the watermarks. Although computers can't tell the watermark from the underlying image when looking at an individual file, an AI trained with thousands of photos using the same watermark is capable of estimating what the watermark looks like on its own.
Google s AI can remove watermarks from images
Google's AI can remove watermarks from images
Google
The AI then gauges the watermark's opacity and uses this information to calculate the actual colours of the pixels beneath it. The result is a rendition of the original photo so accurate humans can't tell the difference.
"A fact that has been overlooked so far is that watermarks are typically added in a consistent manner to many images," Google said of its work. "This consistency can be used to invert the watermarking process – that is, estimate the watermark image and its opacity, and recover the original, watermark-free image underneath. This can all be done automatically, without any user intervention, and by only observing watermarked image collections publicly available online."
The AI is less successful when subtle random distortion is applied
The AI is less successful when subtle random distortion is applied
Google
READ NEXT: Box hooks into Google's cloud to improve image search
The discovery could allow AIs to build catalogues of stock images with the watermarks removed, opening the door to mass piracy. Watermarks can be manually removed using modern image editing software but the procedure is tedious, time-consuming and difficult for unskilled users to complete. AI now offers a way to remove watermarks instantaneously with far greater precision.
Thankfully for stock image providers, there is a simple way to improve watermarks. Google said that applying a randomly generated warp to the watermark each time it's used will prevent the AI from recognising it. The distortion can be "very subtle," leaving the watermark looking almost unchanged by humans, but remain an effective way to defeat the technique.
If every image has a slightly different watermark, the AI will miss the recurring pattern across all the stock photos. Even if it does succeed in identifying the watermark, visual artefacts are likely to remain in the image. However, it may only be a matter of time before newer models are capable of overcoming this solution.