Log in

View Full Version : Neural Net?


cogman
28th January 2010, 22:48
Just something that I've been mulling over, I don't really know how feasible it would be.

One of the harder aspects of encoding is the question "What looks 'good'" The general answer being "it depends.". We have some semi-objective methods of measuring the "Goodness" of a video, but nothing that truly captures what the human eye perceives as good.

So here is my question, would it be possible to train a neural network to measure "goodness" of video? Neural nets are said to be good a pattern recognition ect, so perhaps by feeding into a multi-layered network both the input frame/output frame we could create an AI capable of more accurately spitting out the "goodness" of an encode.

The problem that I see with this is that a neural network is generally not the fastest thing, add in several 1000's of nodes (presumably for the number of pixels you have) multiplied by the number of layers you want to go and you could have a very slow process of identifying how good an image is. So perhaps that may not be the AI answer, but what of other AI algorithms? Perhaps they could be used to produce a better AQ Mode or psy optimization. IDK, just throwing the Idea out there.

Yes, you would have to have a human hand hold it through a couple of videos, but it might just result in some pretty good measurements that people subjectively like.

LoRd_MuldeR
29th January 2010, 03:07
A whole lot of quality metrics have been propose, but the metric that perfectly predicts subjective quality has NOT been found yet.

Heck, even if you take 10 people and let them decide which encode looks best, you'll get 10 different opinions ;)

Anyway, some of the most promising quality metrics have been summarized here:
http://www.ece.uwaterloo.ca/~z70wang/publications/SPM09.pdf

With your idea I would fear that even if you can manage the required processing power, it would depend too much on the video it has been trained with...

cogman
29th January 2010, 04:16
A whole lot of quality metrics have been propose, but the metric that perfectly predicts subjective quality has NOT been found yet.

Heck, even if you take 10 people and let them decide which encode looks best, you'll get 10 different opinions ;)

Anyway, some of the most promising quality metrics have been summarized here:
http://www.ece.uwaterloo.ca/~z70wang/publications/SPM09.pdf

With your idea I would fear that even if you can manage the required processing power, it would depend too much on the video it has been trained with...

How about a user managed training. If the user fed the video and told it what looks good to them, then you might be able to get a good measurement (for certain types of video.)

It is true though, different video types might throw the thing for a loop.

7ekno
29th January 2010, 20:32
It's possible to use multiple users with multiple videos into it's training regime ... an even easier method would be to run a blind study (say 5 samples of 5 videos ranked for "quality" against some clean BluRay sources over 200 observers and an aggregate of 200 video clips as reference) alongside the training studies, then feed both sets of data into the Neural Network (the video clip and the perceived quality score - or even better, the range of scores on the reverse propagation inputs of the Neural Net) ;)

7ek