Machine Learning and Photography

[Timmay]

Photoshop has a new tool called “super resolution.” It uses machine learning to enhance details in the image and make larger resolutions for finer detail and larger prints.
Think of like in the movies where they enhance the crappy security camera image to get a license plate number.
Watch here for a demo if you’re curious.

Anyways some guy mentioned the fact that the machine learning only guesses and it could be wrong on its enhancement of details. It could create details that were never there. It could even create details that a camera lens could never resolve, thus violating quantum physics. He basically says this enhance tool is distorting reality.
The machine learning uses 1,000s of images to learn how to enhance details in ways that reflect what would happen if you actually used a physically higher resolution camera or lens.

So is this feature a distortion of reality/truth, or a tool like a magnifying glass that helps us to see something more clearly?


Sent from my iPhone using Tapatalk

 

[hammondjones]

Well, I think that to some degree, even our own eyesight distorts reality: the brain actively looks for patterns it recognizes, and fills in gaps. E.g., we have a blind spot in each eye, though we typically don't think about it. I don't know that ML is very different from that, to some extent. It takes patterns it sees, and uses that information to fill in the rest, to the degree it is able to.

I think that the bigger issue which will come about is that these tools are quite good at making enhanced images (and videos) look real, and it will become increasingly hard to differentiate between real and enhanced (or faked), at least from a human perspective.