Watermarks easily removed by Google AI

in

An interesting blog post popped up from Google a few days ago. In the post they speak on and demonstrate how an AI they developed can easily remove watermarks from images. It works be taking advantage of the fact that most large companies automate the watermarking process. They create a watermark template that is applied in a uniform manner on all photos. Google’s AI looks at a large collection of images and in doing so is able to built a mask of the original watermark. With an accurate mask of the watermark the AI is able to use that to analyze and remove the watermark from any image that uses it. This graphic bellow illustrates how this work:

Google demonstrates AI that removes watermarks

The frame on the left represents the collection of watermarked images that Google’s AI analyzes. The center image is an composition of all those images combined. You can see as all the watermarked images are merged over time only the watermark remains clearly visible. The image on the right represts what the AI believes the original watermark to have been. The results are pretty impressive and Google is able to show watermarks from some of the leading stock photography websites vanishing off of their images.

However if you are a photographer don’t his the panic button just yet. This is a demonstration, a proof of concept. The software to do this currently does not exist in a commercial package. It will absolutely available one day, but it’s not available today. If someone wants to remove a watermark from your image today they still need to do it the old finished way. With a lot of patience and a lot of cloning in Photoshop.

Also Google goes on to demonstrate how their AI can easily be stopped. If the watermark you apply is slightly warped every time it is applied to a photo then the AI can not accurately remove it and artifacts from the watermark remain on the final image. I guess that gives the people at Google something to work on for version 2.0 of this software.

Check out this video of the full demonstration: