Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Subpixel: A subpixel convolutional neural network implementation with Tensorflow

github.com

Post reply on HN
Print view View original

41–47 of 47 posts

Re: Subpixel: A subpixel convolutional neural network implementation with Tensorflow

#41

Post by taneq »

Eliezer wrote: post #12
maxander wrote: post #2

So, basically, this is the thing in a crime detective movie where the forensic analyst is looking at a terrible pixelated surveillance camera still and says "enhance," and the computer magically increases the resolution to reveal the culprit's face. Just another entry on the "things that are supposed to be impossible that convolutional nets can do now."

And that's how that guy whose face appeared a few times in ImageNet became the world's most wanted terrorist, on the run for thousands of crimes.

I bet he's hiding out with that lab tech whose poor technique lead to their DNA being in hundreds of crime scene samples.

Re: Subpixel: A subpixel convolutional neural network implementation with Tensorflow

#42

Post by RBerenguel »

andrewprock wrote: post #36
utkarshsinha wrote: post #34

Earlier quoted context omitted.

Wait, the neural network encodes within itself probability distributions of the various image patches it has seen. This is sort of like AI. Approaches in the past used heuristics (like finding edges and upsampling them, etc). Those were fragile systems. In this approach, the system learns what's appropriate on its own.

This is not AI in any real sense. It is a fairly straightforward machine learning application to computer vision.

Knowledge is a form of compression

Re: Subpixel: A subpixel convolutional neural network implementation with Tensorflow

#43

Post by joelthelion »

maxander wrote: post #2

So, basically, this is the thing in a crime detective movie where the forensic analyst is looking at a terrible pixelated surveillance camera still and says "enhance," and the computer magically increases the resolution to reveal the culprit's face. Just another entry on the "things that are supposed to be impossible that convolutional nets can do now."

Except it might unblur to the face of someone else.

Re: Subpixel: A subpixel convolutional neural network implementation with Tensorflow

#44

Post by anotheryou »

trustswz wrote: post #24
anotheryou wrote: post #22

Earlier quoted context omitted.

Super wonky indeed. Also it should compare to something like photoshops bicubic enlargement or the original size, because the brain gets stuck on the pixel edges.

If you are interested in how it compares to bicubic or the original. Check these papers using the sub pixel convolutional layer: https://arxiv.org/abs/1609.05158 https://arxiv.org/abs/1609.04802 .

thanks, impressive

Re: Subpixel: A subpixel convolutional neural network implementation with Tensorflow

#45

Post by nerdponx »

witty_username wrote: post #40
andrewprock wrote: post #37

Earlier quoted context omitted.

I would expect AI to include some sort "emergent behavior", so in a sense you are correct. If a program does exactly what we expect it to, exactly how we tell it to, it almost certainly isn't AI. Unless we are telling it to "be intelligent" whatever that means.

> As machines become increasingly capable, facilities once thought to require intelligence are removed from the definition. For example, optical character recognition is no longer perceived as an exemplar of "artificial intelligence" having become a routine technology.[3] Capabilities currently classified as AI include successfully understanding human speech,[4] competing at a high level in strategic game systems (su…

This makes a lot of sense and has the added benefit of forcing us to reconsider what we mean by "intelligence."

Re: Subpixel: A subpixel convolutional neural network implementation with Tensorflow

#46

Post by anotheryou »

anotheryou wrote: post #19

I think it's always problematic to compare to images upscaled via nearest-neighbor. The big pixels are hard to parse for our brain, we detect all the blocky edges. A good content unaware upscaling would be nice (one of the default photoshop algos) I also wonder what they used for the downscaling. I see 4x4 pixel blocks, but also some with 3px or 7px lengths. This looks pixely and is supposed to be a source file?: htt…

from trustswz' comment:

https://arxiv.org/abs/1609.04802

The pic with the boat on page 13 is interesting. In the SRGAN version I would take the shore for some sort of cliff, while the original shows separated boulders.

Re: Subpixel: A subpixel convolutional neural network implementation with Tensorflow

#47

Post by Roboprog »

tree_of_item wrote: post #29
Roboprog wrote: post #4

Interesting image "upscale" algorithm. I'm not familiar enough with the field to understand how the "neutral net" part feeds in, other than to do parallel computation on the x-pos, y-pos, (RGB) color-type-intensity tensor interpolated/weighted into a larger/finer tensor. (linear algebra speak for upscaling my old DVD to HD, that sort of thing) At the risk of exposing my ignorance, this has nothing to do with "AI", ri…

Everything that we understand how to do is "not really AI". It's only "AI" when it's still a mystery. At least that's the way people act.

Fair enough :-)

I missed the part, though, where there was some "learning"/"adjusted predication" in the interpolation function(s), rather than just a fixed calculation such as a literal linear interpolation.

I was happy just to be able to tease apart the big equation before the the python code sample, but was too lazy to drill down into what the "delta-x"/"delta-y" factor-functions were.

Still, this was a good presentation: somebody with little to no knowledge of the field, but some math, could get the gist of it. Kudos to the author.

Post reply on HN