ANN for image comression

[masai]

Did anybody use artificial neural networks for image compression? Is it perspective approach?

Most of articles I found describes only MLP or SOM methods (or mofications) which are well known.

[Matt Mahoney]

Sort of. The mixers in PAQ and ZPAQ are a kind of neural network, and they compress images as well as other data types.

[Ethatron]

I think he wants perceptrons, not statistical networks.

[Matt Mahoney]

Depends what you mean by a neural network. A ZPAQ mixer computes a squashed weighted average of its inputs, then adjusts the weights to reduce the prediction error. That's what a neural network does. Yes, it's not a neural image model because the inputs and outputs are bit predictions, not pixel values.

[Karhunen]

Is this software http://gate.ac.uk/projects/jhu/ i.e. "GATE" allow one to create models for bit predictions, albeit just for text ?

[Matt Mahoney]

Good question. If you can output a probability distribution over word predictions, then you could convert it to bit predictions. The good thing about this approach is that only the low level part of the modeling has to be done for every bit.

[willvarfar]

Somewhat related: http://www.youtube.com/playlist?list...3&feature=plcp http://www.reddit.com/r/programming/...m_image_video/

[Shelwien]

Also
http://clampression.blogspot.com/
http://code.google.com/p/clam-compre...o/updates/list
However, it looks more like a good example of project promotion than anything innovative

[willvarfar]

That's not a problem if he's having fun

Claiming 100x on some of his test images is pretty impressive.