Sharpening suggestion (=deconvolution)

archimedes · Post by **archimedes** » Fri Jul 04, 2003 11:01 pm

I am using NI for some time now, beeing still impressed by the noise removal qualities.

However, I am not as satisfied with the sharpening functions. It looks like the sharpening sliders simply amplify the image amplitude in three selected spatial frequency areas. Sharpening functions in photo editors seem to deliver much more pleasing results, when it comes to compensate high frequency attenuation coused by film, scanners, filter functions and the like. Theoretically there should be a benefit for NI, because it is able to sharpen the image "signal" and not the image "noise", but I cannot see any such benefit.

Since NI seems to do its main task of noise reduction by transforming the image signal into the frequency domain and then back, my suggestion would be to do sharpening by deconvolution, which then would not add much of an additional computational burden. And "noise amplification" - a common disadvantage of deconvolution approaches would not be so critical - since deconvolution could be applied to an aggressively denoised image - like NI-sharpening is done today anyway.

The point spread function could be "trained" by the user in a similar way, as he now trains the noise profile - selecting certain areas. However, these would not be completely featureless rectangular areas, but straight lines in the image, which the user knows to go across a single "intensity step function". E.g. if there is a house in a scene with a locally uniformly red color against the plain sky of a uniformly blue color, the user could draw a straight line from the inside of the roof to the sky - perpendicular to the top of the roof. Ideally the intensity along that line would form a perfect step function from roof color to sky color. In reality, the step function is smoothed out due to film, optics and scanning "softness". The "smoothed function" is nothing else than a one dimensional projection of the point spread function for this particular location and direction in the image. By selecting several of such training areas, which the user wants to have sharpened, he could NI have to build a complete representation of the point spread functions for different parts of the image. Deconvolution of the image using the collected PSFs could the follow the textbook approach.

Just a suggestion

Stefan

Post by **NITeam** » Sat Jul 05, 2003 9:06 am

Yes, there is some potential in that image restoration approach based on PSF as described in the textbooks. However, there are many technical obstacles to practical use of this approach starting from more complex device mode models to more complicated training procedures. I suspect we see the lack of such tools exactly because of these obstacles. Generally, it is a whole new task not very much related to noise reduction and it should be solved in a separate software application, which will probably be one of our future development steps.

Thank you for the suggestion,
Vlad