Google’s new AI-powered, super-resolution image technology certainly looks impressive, but what are the implications for a digital imaging industry that has historically used downsampling as a form of protection against theft?
While many of us still regard artificial intelligence (AI) as something out of science fiction movies, its role in digital imaging has made it more a part of our everyday lives than we might think. And its usage is only set to increase.
If you have an interest in photography, you might be familiar with the term in relation to your camera and the editing software you use, but beyond this, there is a huge number of other applications for the technology. These include the moderation of social media content, medical diagnostics, and driverless cars, among many others.
However, one of the most recent breakthroughs in the use of AI in imaging is Google’s new image upscaling (or super-resolution) technology, which is designed to increase image resolution.
Announced in a blog post by Google AI – a division of the tech giant dedicated to artificial intelligence – it is called Super-Resolution via Repeated Refinements (SR3) and uses deep learning, an advanced form of machine learning that is based on artificial neural networks.
In this article we provide an overview of the technology and discuss how, while designed for good, there could also be a darker side.
What makes Google’s AI-powered super-resolution imaging so special?
AI-powered image upscaling is no new thing; the technology has been around for some time, courtesy of a raft of online apps. But while the results from some of these have been impressive, traditional models have struggled to produce perfect results, with upscaled images often appearing blurry.
Google has taken a new approach by adopting a different type of deep generative model called a diffusion model, which it believes is the more stable and higher-quality option.
A diffusion model works by taking a high-resolution image and gradually adding Gaussian noise until the image details are obscured. The process is then reversed, slowly ‘de-noising’ the image, adding details back in until it reaches full resolution.
By doing this many times, with many different photographs of many different subjects, it is possible to develop an optimization algorithm for the process.
Using this model, SR3 reduces a low-resolution input image down to pure noise, then regenerates it as outlined above. However, through extensive training on countless images, SR3 is apparently able to predict the most likely pixels required for it to continue adding detail above and beyond the input image’s original resolution.
While some small imperfections can be seen, the results appear remarkable. And when pitted against other face super-resolution methods in a two-alternative, forced-choice experiment, Google’s technology was the clear winner, with 47.4% of respondents choosing the SR3 image as the genuine version. See it for yourself here.
The problem with image upscaling
The benefits of this new and improved deep-generative image upscaling are both undeniable and abundant. For example, when futureproofing older images that were taken on a device with lower capabilities or optimized for outdated screen resolutions, this standard of upscaling presents a fantastic opportunity.
And that’s just on the most trivial level. With the potential for more efficient medical diagnostics and improved safety of driverless cars, there is undoubtedly a very important place for this technology in the future of our everyday lives.
However, while bringing much good to the world, what will this powerful super-resolution imaging mean for the security of digital images online?
Consider this: If you own the copyright to a particular image, in the vast majority of cases you would have access to the original file at maximum resolution. So why would you need to upscale it?
Furthermore, when you consider that over 75% of the world’s population carries a smartphone in their pocket, the majority of which possessing the power to create more than enough resolution to meet a user’s requirements (and some soon to push it to the extreme), it’s clear that the everyday user would have very little legitimate use for this technology.
In fact, many photographers deliberately downsample their images before sharing them online as a form of security against theft, the logic being that, while the images can still be stolen, their low resolution significantly restricts options for misuse.
Therefore, these advances in image upscaling could potentially expose trillions of downsampled images to fraud, affecting countless livelihoods as a result.
How can we benefit from image upscaling while ensuring complete image protection?
This is certainly a concern for many artists, photographers, and content owners around the world. With an industry reportedly already facing up to €532.5bn of annual losses through the theft of digital images, could this be the nail in the coffin?
There are other means of protecting images online, such as watermarking, adding metadata, and using Google’s reverse image search to detect unauthorized usage. However, these methods – like the practice of downsampling – do nothing to tackle the root of the problem: image theft.
Image streaming, on the other hand, provides complete image protection online. The technology works by a user storing their images in a secure central bank and streaming them to websites using an embed code – much like videos are streamed on YouTube.
Not only are the streamed images protected against dragging and dropping and right-click actions, but the image owner can also see a list of URLs that each image appears on, with the ability to block websites as they see fit.
What’s more, with Hyper Zoom functionality and full-screen viewing providing maximum detail, while maintaining fast load times (as illustrated in the SmartFrame below), image streaming also eliminates any need for photographers to downsample their images.
Conclusion
Until SR3 becomes available for testing, it remains to be seen just how effective the technology is. However, judging from the results published by Google, it appears to be a huge step forward in both imaging and AI that will bring outstanding good to the world.
That said, in the wrong hands, it has the potential to be devastating for the imaging industry. So, the question of how widely the technology will be available once released is sure be on the lips of photographers and rights-holders the world over.
With so many potential implications, it’s never been more important to ensure all online images are protected from day one. That way, the industry can simply enjoy the benefits of SR3 and other amazing forms of deep generative super-resolution imaging, without worrying about the potential harm this technology could cause when used by bad actors.
To learn more about how SmartFrame image-streaming technology can help you protect online images while making you money through in-image advertising, get in touch with our friendly team today.
Subscribe to our newsletter
Get the latest articles and updates delivered straight to your inbox