Find the difference between images

Dear, masters in matlab & neural networks, sorry for my English. Please, advise me if it is real to develop neural network that will compare two images(original and its compressed version) and return the distortion level between them? If 'yes', what kind of network should be used? May be, someone has examples?

[Image Analyst]

To be clear, you mean with the compressed version once it's been decompressed.

[Walter Roberson]

Rotated? Translated? Cropped? Resized? Or _exact_ size and image position matches?

Answer 1

[Geoff]

Why wouldn't you just subtract one from the other and use some statistics like mean, variance, etc?

Answer 2

[Image Analyst]

PSNR http://en.wikipedia.org/wiki/PSNR is often (usually?) used for that. You might also look at Stuctural Similarity (SSIM) http://en.wikipedia.org/wiki/Structural_similarity

Answer 3

[Greg Heath]

You said that you have found a reference but have no access.

An obvious way to begin is either obtain access to the reference or obtain access to one of the authors.

To use a neural net you have to train it with typical examples of input-output vector pairs.

From what I've read so far your input is a 64-dimensional input vector obtained from columnizing an 8x8 window of a difference image and your output is a scalar measure of similarity.

The enigma here is how to calculate the MOS to use for training.

Once that is defined, you don't need the network.

Or am I missing something?

Hope this helps.

P.S. Use windows with odd numbers of pixels per edge so that the middle of the window is at a pixel location.

[belka0011]

Yes, an input is 64 - discrete cosine transformed pixels.

I think that I should use sigmoid function and will try it in nn toolbox

Answer 4

[belka0011]

I want to develop NN that will better in searching artifacts than PSNR.

I've found description of work with almost similar problem, but I haven't access there((

If be anyone can help me?

[Walter Roberson]

What kind of artifacts are you referring to for your purposes?

Are you talking about weighting some differences as more important based upon surrounding context?

[Image Analyst]

Better how? What's "wrong" with PSNR? What shortcomings do you feel that it has? PSNR is a well established technique that gives a single number that is used as a quality metric for comparing an original image with one that has undergone the compression/decompression round trip. Your method will likewise give a single number. Do you plan on doing psychophysical experiments to prove your number correlates better with human perception than PSNR? Maybe you ought to read this paper first: http://spiedigitallibrary.org/jei/resource/1/jeime5/v19/i4/p043004_s1?isAuthorized=no

[belka0011]

thanks for the reference...but I should use neural network and find artifacts after embedding digital watermarks in image

[Walter Roberson]

So there isn't any compression step?

[Image Analyst]

belka, but you still didn't answer my questions.

[belka0011]

2walter: no compression, only embedding watermarks

[belka0011]

2image analyst: "Your method will likewise give a single number" yes.

And, yes, I will do psychophysical experiments

[Image Analyst]

And what shortcomings do the current methods have that your method will overcome?

[Walter Roberson]

http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.5.6925&rep=rep1&type=pdf

Answer 5

[belka0011]

I cannot say why my idea better than PSNR, that's my scientific work...

Now I've decided to develop neural network, that will contain pairs: artifact(8x8 pixels in DCT) and coefficient(from 0 to 1, depending on it visualization).

And now new question, how better to realize it?

[Image Analyst]

Better than what? What do you have so far? I'd say the Neural Net Toolbox would be a good thing to help you realize it.

[belka0011]

Better for this aims, what performance function and what training function to use?

[Image Analyst]

I don't use neural nets. Sorry - I can't help.

[belka0011]

anyway, thanks for your answers =)