Watermarks might not be the best way to prevent people from stealing stock photos online. At least, that’s what a Google research paper presented at the 2017 Computer Vision and Pattern Recognition Conference (CVPR 2017) suggests.
While it’s difficult for a computer to remove the watermark from an individual image — and while it can take Photoshop experts quite some time to manually remove watermarks — a computer can analyze numerous watermarked photos and estimate what the image beneath the watermark looks like.
“The first step of this process is identifying which image structures are repeating in the collection,” reads an excerpt from an August 17th, 2017 Google Research Blog post. “If a similar watermark is embedded in many images, the watermark becomes the signal in the collection and the images become the noise, and simple image operations can be used to pull out a rough estimation of the watermark pattern.”
The whole process works by analzying a substantial number of stock photos, in order to determine the approximate position where their watermarks are typically located. Then the watermarks themselves are categorized, based on image and opacity components.
By doing so, a computer can effectively separate a watermark from its base image, and reconstruct a new image without a watermark.
In a word, the results are impressive.
The paper provided several examples of stock images with their watermarks removed, and while MobileSyrup hasn’t done any extensive graphical examination, it’s difficult to tell that these images once had watermarks with the naked eye.
The paper also outlines a number of steps that stock photography sources can take to better prevent the removal of their watermarks.
“We need to introduce inconsistencies when embedding the watermark in each image,” reads another excerpt.
These inconsistencies include warping the watermark’s text and image, to prevent computers from accurately categorizing a watermark’s image or opacity components.
What results is an edited photo that leaves behind clearly visible evidence that the image once contained a watermark.
“In a nutshell, the reason this works is because removing the randomly warped watermark from any single image requires to additionally estimate the warp field that was applied to the watermark for that image — a task that is inherently more difficult,” reads another excerpt.
While this paper doesn’t necessarily speak to any present or future Google products, it’s difficult not to draw similarities between this watermark removal process and the photo editing tech that Google unveiled at their I/O conference in June 2017.
The photo editing technology enabled users to remove unwanted objects from a photo — like a fence obscuring an image of a child playing baseball — without needing professional tools.