Image super-resolution through deep learning – GitHub | Hacker News Reader

Image super-resolution through deep learning – GitHub

github.com

53 pointsdavid-gpu14 comments

Author here -- first time submitting anything to HN. Please let me know if you have any suggestions.

would love to see some non-faces examples

It's on the way. That particular network was trained with a face dataset.

Fantastic result!

It would be valuable to see statistics about how well this algorithm does against the face database itself. De-rez a sample of images from the database, up-rez them and then find them in the database using face recognition. How often is the original image identified from the up-rezzed version?

It would also be interesting to see how different the up-rez results are if the original is cropped differently, shifted, slightly rotated, darkened or lightened.

Suuuper late to the discussion, but there is a project for anime-style images and denoising that uses CNNs.

https://github.com/nagadomi/waifu2x

Have you tried using it on surveillance photos (such as police requesting citizens help). These are often of poor quality.

I have not done that yet, but that was basically my original motivation. The main difficulty would be obtaining a good dataset for training, specifically a dataset where most of the faces are not looking straight ahead at the camera, and where the camera is located above eye level as would be the case in a typical security camera.

from movies and classifier that looks through images.

Holy shit, Enhance is actually possible!

But only usable if you want to catch the wrong person.

Next step: Require photos from multiple angles of all your citizens. Feed into database.

Impressive! Would be great to see more examples.

It would be more impressive to see the variety of features that could be the image in question. 16 possibilities if I am correct.

Sorry, I don't understand. What sort of features are you thinking about? And why only 16 possibilities? There's a nearly infinite number of ways in which the average of 4x4 pixels can be a particular value.

Reply on news.ycombinator.com