PDA

View Full Version : Super-Resolution



Nuradnan
2006-Jan-04, 07:12 PM
I read a paper about the enhancement of small telescope images using Super-Resolution techniques. I have some questions related to this.

First, is the High-Resolution (HR) result image is real? I mean, the HR is produced from statistical process. There must be some uncertainty involved in the process. For instance, if we see blur Low-Resolution (LR) images of Mars, and then see pattern representing canyon in the HR image (in extreme case), is the pattern really canyon? Isn't it merely some pixels containing low intensity due to statistical process? How much the uncertainty level then?

Second, what is the maximum increase of the spatial resolution? Well, there is a joke between I and friends that with small portable telescope and Super-Res technique we can see the Mars rovers wandering .. :)

Third, what is the basic algorithm of Super-Res? I've read some methods such as RE-ST, LAZA, SIAD, DAEI, and bicubic interpolation. Though I know the very basic idea about Super-Res, I still don't know the details. What's the difference between Super-Res for single-frame-image and multi-frame-image? Would you like to suggest some reference papers and/or books to read? I really want to know the concept and technical things from scratch.

Phew. Thank you very much .. :)

ngc3314
2006-Jan-04, 08:29 PM
I read a paper about the enhancement of small telescope images using Super-Resolution techniques. I have some questions related to this.

First, is the High-Resolution (HR) result image is real? I mean, the HR is produced from statistical process. There must be some uncertainty involved in the process. For instance, if we see blur Low-Resolution (LR) images of Mars, and then see pattern representing canyon in the HR image (in extreme case), is the pattern really canyon? Isn't it merely some pixels containing low intensity due to statistical process? How much the uncertainty level then?

Second, what is the maximum increase of the spatial resolution? Well, there is a joke between I and friends that with small portable telescope and Super-Res technique we can see the Mars rovers wandering .. :)

Third, what is the basic algorithm of Super-Res? I've read some methods such as RE-ST, LAZA, SIAD, DAEI, and bicubic interpolation. Though I know the very basic idea about Super-Res, I still don't know the details. What's the difference between Super-Res for single-frame-image and multi-frame-image? Would you like to suggest some reference papers and/or books to read? I really want to know the concept and technical things from scratch.

Phew. Thank you very much .. :)

I replied to a similar query as follows some months back (and this cut-and-pasting avoided having to hunt the reference again):

There is a technical review, if you're looking for the mathematical details, available here. A lot of the details of implementation are still sort of the laboratory art of astronomers and not widely written up outside of specific applications. For example, it is a sort of superresolution to compare a well-sampled image of a suspected binary star to a known single star, and ask whether the pixel values are better fit by one or two (slightly offset) components. You can also google "drizzle processing" for a more general approach, improving image sampling by combination of multiple offset images of the same field (as done in the assorted Hubble Deep Fields and now a part of the Hubble Advanced Camera pipeline processing).

The diffraction limit is not a sudden cutoff - there is still some information in an image beyond the diffraction "limit", but progressively more attenuated on finer and finer scales. In superresolution, one is trying to use what's left of this information without being overwhelmed by the noise, which becomes more and more dominant toward higher spatial frequencies. This is less overwhelming in radio interferometry because the noise does not occur directly in the image, which is why radio astronomers have been happily doing deconvolution for forty years now and it remains a ticklish process for optical images.

-----------------------------------------------------------------------

New text: several algorithms are in use. For very high signal-to-noise and accurate blur functions, Fourier processing can yield information right up t the sampling of your pixels (since convolution in the image domain is multiplication in the Fourier domain). For more usual cases, knowledge of the noise level is key - one popular algorithm, maximum-entropy, generates the smoothest image which is statistically consistent with the data and noise model. On top of the noise and other data-quality issues, the more information you have at the outset (for example, a blur pattern is independently known to consist of three unresolved stars seen close together) the more reliable the results.