How to correctly determine which of the images contains less noise?

esc2012-08-09 15:59:33

Algorithms

esc, 2012-08-09 15:59:33

I compare the same images, but with different quality (taken in different conditions, reaped with different losses). I want to calculate the best quality.
For example: The contrast and the number of colors are higher at the bottom (thanks to the noise). Can someone suggest more intelligent algorithms? (there is no standard, the initial data are only such).

Answer the question

In order to leave comments, you need to log in

10 answer(s)

shipovalov, 2012-08-09
@shipovalov

in the eyes of a non-professional, the top looks better

Maxim Kuzovlev, 2012-08-09
@KY3EH

I would recommend looking in the direction of noise reduction algorithms, perhaps there are metrics for determining the noise level of an image. For example, here is this article .

Infthi, 2012-08-09
@Infthi

well, I would try to find the difference between pictures and pictures processed with noise suppressors (a primitive version is blur) - the difference will be larger for a noisy one.

Maximus5, 2012-08-09
@Maximus5

It's all relative. See what the task is. After all, the top one has a much smaller dynamic range, and if you stretch it (for example, AutoLevels), you can get the same garbage as the bottom one.
If you figure out some general criteria (after all, they depend on the task), then you probably need to look at the histogram. I suspect that the one in which it looks more like a “bell” and is located closer to the center (such as overexposure / underexposure) will be better. In theory, the smoother it is (without “combs”), the better the picture.

impass, 2012-08-09
@impass

Google "image quality assessment" if you speak English

Akson87, 2012-08-09
@Akson87

You first determine what you think is the best picture. People argue about the quality of pictures on all sorts of TVs from different players, etc., etc. (and you can also remember audiophiles with a warm tube sound :)). That's when you determine what is good and what is bad for this case, then you can already think about the algorithm.
I can assume that if you run them through FFT and see which high-frequency part is smaller, then there will be less noise, although everything can simply be hard-wired there.

Andrew, 2012-08-09
@xaoc80

You can also do this - binarize the image (for example, by average brightness) and where there are fewer black dots, the quality will be higher

Wott, 2012-08-09
@Wott

noise will spoil everything, so first you need to remove them,
then compress and see the percentage of losses,
the quality of the shooting can be determined by the width of the histogram

Eddy_Em, 2012-08-10
@Eddy_Em

I would try to build a histogram of the HL + LH component of the wavelet images of images (by Haar, for example). True, it will not be very objective.
You can also smooth the pictures by the median (say, 3x3 or 9x9, depending on the size of the original images) and subtract from the originals. The STD of the residual may well be a true characterization of the noise.

Infthi, 2012-08-12
@Infthi

By the way, I just came up with another idea. Judging by the pictures, are you trying to compare the quality of the films? :) And in films there is usually such a thing as credits - a black screen with high-contrast white characters. It seems to me that comparing the quality of screenshots of titles will be much easier.