Monday, September 25, 2017

Protecting your images from automated watermark removal

Image showing strong watermark with unique Image IDMany who market or sell photographs online or who just want to prevent online photo theft of the images they've posted for family and friends, often use watermarks as their first line of defense. While it's not the only defensive measure photographers can and should take to protect their images online, watermark use is sensible, practical and useful.

Although its true that someone with expert Photoshop skills can eliminate a watermark in an hour or so, even if thoughtfully constructed, watermarks still stop most thieves because it's rarely worth spending an hour or more to steal an image, especially if the final product has any telltale visual artifacts shouting, “Theft!”

This past July, at the Computer Vision and Pattern Recognition Conference, Google demonstrated an algorithm capable of automating the removal of watermarks from photos. Google showed that the removal of watermarks from a series of photographs that now takes hours or days, could be done in minutes.

Google's new algorithm can identify the recurring visual patterns found in watermarks, even in a wide variety of watermark types, including those used by Shutterstock, Getty and Adobe, then automatically strip them from images with no direct input from a human editor. In the demonstration, a computer was able to remove the watermarks from several hundred images in just a few minutes.

Google was at the conference to explain that the algorithm, developed from machine learning and artificial intelligence research, wasn't created to damage the photography and graphic arts industry. In fact, Google specifically demonstrated the algorithm to let the industry know it existed and make public knowledge about how it worked, to be able to defend against it, if others developed similar algorithms for sinister purposes.

To use it, a user scans, analyzes, and catalogs the patterns within the watermarks on about a 1,000 images using a computer running the algorithm. The computer identifies everything about them visually, including the gradients, opacity, shadows, and geometries. That enables the algorithm to isolate each watermark from the image behind it, erase it and then fill in the blank pixels, from where the watermark was removed, by extrapolating the remaining image around them.

At the conference, Google explained that the algorithm essentially works because companies and individuals in the industry typically repeat the same basic watermark over and over again on their images and graphics. That reason for it working so well, it turns out, is also its “Achilles Heel.”

If each watermark is somewhat different, the algorithm can't totally isolate the watermarks from their images and that in turn means when the watermarks are removed the algorithm leaves easily recognizable telltale visual artifacts.

A more detailed analysis of the algorithm's shortcomings also reveals that small changes in the opacity of watermarks doesn't minimize the algorithm's ability to successfully remove them, but highly opaque watermarks are more difficult to remove. In addition, merely repositioning watermarks also doesn't reduce the algorithm's success, unless the watermarks are moved from a smooth or regularly patterned area of the images, to an area that is more complex. When a watermark is over a complex area of a photograph, it's more difficult to not leave visual artifacts when it's removed.

To combat the algorithm, Shutterstock now randomly alters their base watermark shape slightly for each image and adds the contributors name to it, to differentiate them and make them more complex.

Getty uses watermarks that are large and each one now includes the photographer's name and a unique number for the image.

For individual photographers posting images online, modifying the size and tweaking the shape of their watermark from image to image may be too difficult to handle. Adding the name of the photographer won't help because it will be the same each time.

I suggest that individual photographers consider the following actions to protect their images with watermarks to make them stronger:
  • Add a unique image ID to each watermark for each image.
  • Change the position of the image ID in juxtaposition to the remainder of the watermark.
  • Place the watermark over the most complex area of the image, when possible.
  • Make the watermark as opaque as possible.
  • Make the watermark as large as possible.
Each of these actions makes it less likely that watermark removals by either the Google algorithm, in the future, or by a Photoshop expert today, or in the future, will be clean and not leave obvious visual artifacts behind. Therefore, I suggest photographers should implement the above actions today to protect themselves immediately.

In addition, to protect your online images consider using the following specifications for your images:
  • Keep the longest side of each image to no more than 800–1,000 pixels.
  • Reduce image resolution to no more than 72–125 pixels per inch.
  • Ensure your copyright notice, rights usage terms, name, and contact information are in your image's metadata.
Coupling those specifications with a strong watermark can help protect your images from online theft.

Finally don't forget to register your copyright. It's essential for your long term protection.


Harvey-Miami said...

Wow, it gets harder and more technical every day to protect one's work. Great article. Thanks.

Denise said...

I don't want to make the watermark too big or it makes the image unsalable. Does it matter where I get the image ID?

Ned S. Levi said...

Denise, when using watermarks, we all have to balance copyright protection with the ability to sell of our images if the size and position of the watermark obscures too much an image and/or a critical part of it.

That's why visible watermarks can't be our only line of defense protecting our images.

As to the image ID, the important thing is that each image's ID is unique and that you keep track of them for later identification if necessary. My images are protected by a unique registered Image ID embedded in an invisible watermark on image which is coupled with my Digimarc ID. The unique image ID shown in my visible watermark on the face of my images is that same Digimarc Image ID. That way I'm consistent and have multiple ways to track the ID and images. I keep a database of the Image IDs.

Denise said...

Thanks Ned.

Will-Lexington said...

Will the simple of an Image ID really make a difference? It's such a small difference.

Ned S. Levi said...

Actually it will. What that does is cause the algorithm to not have the ability to make a super-clean removal of the watermark. There will be telltale visible artifacts left on the final image after the watermark is removed that should be easy to see.

There's nothing that can be done to prevent someone using automation or to manually remove a watermark. The preventative is to make it hard to remove cleanly to the illegal removal can be detected and to make it take a very long time to remove it, so many people won't think it's worth the effort.

So, yes it's makes a difference, but it's not foolproof, which is why I suggest other anti-theft methods and registering your copyrights.

Pete - St. Pete said...

I've started to implement your watermark strategy tonight by adding an image ID to every watermark. Thanks for the information and tips.

Debra-Los Angeles said...

Ned, I see you have implemented the suggestions in your article, but is there anything else we can do to prevent online theft of our images? Is there anything else you can do?

Ned S. Levi said...

Debra, in addition to what I wrote in the article I use an invisible Digimarc watermark. This doesn't prevent image theft, but it does help to find out if one of my images has been illegally used, where it's displayed online, and to prove it's my image. I highly recommend the Digimarc product.

Denise said...

I'm following your advice and have added a unique image number to each of my watermarks.

Post a Comment