Bigger images, better tools control?

[papaf]

Hi all,
since I read a thread about a way to do drizzle with pixinsight, I tried it on a couple of images. My camera, an Atik 314L, normally produces small images, 1.4Mpixels only. With drizzle, the image starts to get bigger.
What I noticed is that many tools, namely ATrousWavelet and noise removing tools are more manageable. I can swing with values more freely, and some big changes are needed in order to see the results.
Actually, I'm stunned by the program changes, knowing the original data is always the same.
Is something like this to be expected? What could be the reason?

Fabio

[Simon Hicks]

Fabio,

I usually use a DSLR with relatively big images....but I have noticed the effects you mention when resizing up much smaller images. So I think what you are seeing is expected. And if I'm illustrating some processing feature or tip to someone who has posted a tiny jpeg image, its usually best to resize it up a bit before processing it....you do get better results.

And I recently took some 10Mpixel type images and resized them up for some posters. The stars looked much smoother....the resizing algorithm actually improved things at the smaller scales....at least to my eye.

Regards
           Simon

[Juan Conejero]

Hi Fabio,

Good question. Bigger images have advantages and drawbacks. Two well known drawbacks are increased memory requirements and slower process execution. Both problems grow, in general, as the total surface in pixels, so they tend to worsen quadratically at least.

One of the advantages is a larger and hence more accurate modeling of the PSF. This can be particularly important for wide field images, and is beneficial for several algorithms and tools, especially for those that perform image restoration and sharpening tasks. For example, if the PSF has to be represented as a 5x5 or 7x7 kernel, then we may have a poor discretization and hence poor/uncontrollable deconvolution results. By upsampling the image, we can use discretizations over 9x9 or 13x13 kernels, which are much more accurate. Upsampling is usually done with the IntegerResample tool in these cases. Note that this not only affects deconvolution, but also wavelet-based sharpening, as in this case we can work with larger, more accurate scaling functions that yield better and more controllable results.