Dear Amit,

yes, of course modern filter technology can help to reduce noise and to keep the high-frequency components of an image. Please have a look to the papers of:

Kostandin Dabov, Alessandro Foi, Vladimir Katkovnik , Karen, Egiazarian.

You will find them here on Researchgate. in addition, the following website isi very good:

http://www.cs.tut.fi/~foi/

I hope that this information will help you. 

Regards Reiner

Dear Reiner,

Thank you for your answer. I wasn't sure if these new methods such as BM3D can preserve high-frequency components of the image. I am going to use the sub-pixel image registration to find very small shifts (deformations) between two images, so I was uncertain if using these algorithms would cause changes on the real data which affect the sub-pixel shift estimation process. 

I will try them and will let you know about the result.

Thank you very much for your help.

have you tried wavelet denoising? Take a look at the following link, hope it helps

https://www.ceremade.dauphine.fr/~peyre/matlab/wavelets/content.html#14

Please, look arXiv:1206.3594 [cs.CV](xxx.lanl.gov). Noise can be considering as blur. So, the filter should be such as deblurring filter

Abusctropic diffusion is also can be useful for removing noise without removing most part of the image. 

http://en.wikipedia.org/wiki/Anisotropic_diffusion

Hi, I am also working on denoising the impulse noise. You can check out paper in my account titled 'Efficient digital image denoising for gray scale images'. It is published in Springer with SCI indexing with impact factor more than 1. I am still working on the same and it is improved again.

Dear All,

Thank you very much for your suggestions. To keep you updated, I checked BM3D, deblurring filters and wiener filter for the images I have. They make images look smoother but also caused considerable decrease in the accuracy of the next step of my algorithm for finding the sub-pixel shift between two images.

It means that both of these algorithms have removed part of the intensity information (especially high-frequency components) which I have used for sub-pixel measurements. I suspect that wavelet filters may cause the same issue.

I will keep you updated.

Thank you

You make an assumption that high frequency information is the most important part for finding accurate sub-pixel positions. Actually, whether this is true will depend on the exact nature of the algorithm you use to calculate the position, and it is perfectly possible to compute very accurate sub-pixel shift from very soft images.

What they look like to you is entirely irrelevant here. The proof of the pudding is in the eating, and if ordinary low-pass noise reduction gives you improved final results in terms of position, I, for one, would not be surprised at all. The more fancy noise-reduction techniques will introduce more intractable non-linear adaptive changes, which are more likely to interfere with your position estimation in unpredictable ways.

This suggestion  is just a conjecture . but more experience is certainly available in previous comments. Open to correction  and apologies for a different unproven  view.

1.Perhaps one should try thresholding  the image first so that as much as possible

of constant value images are removed. Even a band type slice with average and +-

values could be tried till we obtain  a reasonable an acceptable slice. We could then filter out noise as those having max values. a measure to determine the best slice would perhaps be minimum max points.

2. A naive Bayes classifier could then be used to isolate further impulse noise points.

3. The resulting slice may contain high frequency needed information which may be

processed as needed.

Good luck

 Cheers

Thank you all for your suggestions. 

Dear Narasim,

Your suggestion looks interesting! I had a similar idea in my mind to try machine learning techniques to detect noise in a small neighbourhood.  What may make that difficult is the low-level of noise that my images have. 

I will try that out and will let you know about the result.

Thank you.

Cheers

Dear Amir Hajirassouliha 

Yes machine learning and probability seem attractive answers. Viewed as an element in a matrix of pixel intensities how does one break this value into components which have some relation to other elements( something like perfect reconstruction in multirate systems).It would seem prior knowledge of the object is one tool.But there is another question Is it possible to discover unkown patterns without this prior knowledge given a set of numbers in say 2dim ?The subject would appear to have many other implications ( pardon the diversion) ... viz Fractals in data,, Big data analytic s  and converging on to IBM's Watson..A very different approach on images is the somewhat difficult  but nevertheless fascinating view in the thread here in RG :

John Jupe asked a question in Instrumentation:

Has the invention of the camera been biggest single roadblock to our developing understanding and the progress of knowledge?

Wish you success

Cheers

Old denoising techniques consider noise as high frequency. But images may have high frequency components that represent rapid changes of the image. Many adaptive filter has been proposed that removes noise as well as preserve high frequency components. Anisotrophic diffusion, Non-local means are such filters for noise removal.