I first reported this last week on TheInsideTips.com, but this is so important, I want to share it with everyone this week.
For years, editors have used mosaic and blur effects to hide the identity of on-screen talent. However, recent research has found a way to reverse-engineer a high-quality image of the speaker’s face from a low-resolution blur. Here’s what you need to know.
Research published in Sept. 2018, from universities in the US and China has revealed a technology that “learns to reconstruct realistic [image] results with clear structures and fine details.”
Using a low-resolution image (on the left), their technology creates a high-quality result using off-the-shelf computer hardware and nVidia GPUs. The researchers discovered an algorithm, powered by AI, “to directly restore a clear high-resolution image from a blurry low-resolution input.”
“Extensive experiments demonstrate that our method performs favorably against the state-of-the-art methods on both synthetic and real-world images at a lower computational cost.”
Here’s a link to their scientific paper. The text is highly technical, but the images are frightening, if you are a producer charged with protecting someone’s identity.
KEY TAKEAWAY
If you want to protect the identity of an on-camera speaker, don’t shoot their face. Today’s technology makes blurs, mosaics and low-res images completely ineffective.
13 Responses to Blurs No Longer Protect Privacy
Any idea if a solid color block over a face would prove more effective? Seems like it would.
My first notion would be to track a solid shape or even a fake face image over the original, then add blur to that.
Mark:
Anything that replaces the source pixels would work, I think.
Larry
Ron:
That would work.
Larry
Larry, thank you for this extremely valuable insight!
Oh brother (BIG Brother, that is)…
Time for me to get into hotel/motel management.
This algorithm would be a godsend for bringing slightly defocused actors back into focus when working with low DoF and they are slightly off the focal plane, particularly in larger ensembles.
Martyn:
I agree – I was almost fired for shooting a senior vice-president somewhat out-of-focus. But the disadvantages of this research are also serious and significant.
Larry
It seems to me that with this technology, couldn’t they turn standard def video into high def by increasing the lines of resolution and applying their pixel magic?
Steve:
It’s possible, but that wasn’t the focus of their research. Their research paper explains more about their goals. However, that doesn’t mean someone else couldn’t invent a higher-quality SD scaler.
Larry
LG’s new TV sets launched today feature AI to enhance faces and text, probably because they know they won’t be getting real 10-bit 8k anytime soon.
Not 100% convinced at all.
Problem with the actual scientific paper is that they never compare their reconstructed faces result with the original…
And their intermediary steps feel somewhat wrong.
Except if their AI thing simply search for a match on the internet like TinEye.
If so it is almost bogus.
Sounds too good to be true for it would be indeed great for many application in our trade, indeed like turning SD into HD – much easier than what they appear to achieve.
Sometimes scientists are overenthusiastic… or preemptive of the field to get attention.
I would wait for more details before declaring LOW RES blurring outdated.
HIGH RES is basically solved, and consider image reconstitution by reflection off a wall (https://www.nature.com/articles/d41586-019-00267-x), but low res destroy info.
Roger:
Thanks for your comments. While I don’t understand all the science in their paper, I did not get the sense they were simply searching the web for a “close-match image.”
However, for me, the much bigger issue is that technology is rapidly advancing. And, if this team has not fully solved the problem, another one will in short order. Since films last for decades, we need to change our thinking regarding how we protect the identities of people on camera. That, to me, is the bigger issue I wanted to address.
Larry