Drizzle Image Weights

[LarryC]

After I drizzle integrate my images PI gives me a "Image Weights" image.  What is this and what can it be used for?

[Juan Conejero]

Hi Larry,

Each pixel in a drizzle integration weights image is proportional to the weighted pixel coverage in the integrated image at the corresponding coordinates. Yes, this looks pretty complex

The following figure represents the projection phase of the drizzle integration algorithm:



The blue squares are the drops extracted from one of the input images. They are smaller than the original pixels by the drop shrink factor. The drops are projected over the output integrated image by applying a geometric trasformation previously computed by the StarAlignment process. As you can see, each drop may cover one or more output pixels partially. We call pixel coverage the total sum of drop fractions that contribute to an output pixel value, in square input pixel units. In addition, each drop is multiplied by the statistical weight computed for its input image by the ImageIntegration process. Finally, rejected pixels are assigned a weight of zero.

As usual in PixInsight, drizzle weights images are normalized to the [0,1] range, where zero means no pixel coverage and one means maximum coverage. So if a pixel in the drizzle weights image is white (one), that means that the integrated pixel has received the maximum coverage in the drizzle integration process. The maximum coverage is always reported on the console at the end of the process. If a weights image pixel were black (zero) it would denote a dry pixel, that is, the corresponding pixel in the output image wouldn't have received any contribution from any input image. While dry pixels are unlikely in a normal drizzle integration, they can happen with few input image sets and very small drop shrink factors, and/or extreme pixel rejection.

Drizzle weights images allow you to evaluate and compare the overall quality of your drizzle integration processes. Ideally, you should get drizzle weights images as uniform as possible. They will depend mainly on the amount of integrated images, the quality of the dithering performed at the telescope, the amount of rejected pixels, the similarity of input images in terms of SNR, and the drop shrink and drizzle scale parameters.

[jkmorse]

Juan,

I would like to get your thoughts on whether there are any benefits from using drizzle for images that are right on the cusp of being well sampled.  I know the benefits in resolution when using drizzle with undersampled data but is there any upside in your opinion if you are working with large detailed data sets (40 subs at either 600 or 900 seconds) where the arcsec/pixel to FWHM ratio is right around 3 (in my case my resolution is 0.73 arcsec per pixel and I am able to shoot FWHM's with AO as low as 2.0).

Is the possible gain so small that it isn't worth the cost of the enlarged master image?  And is there any downside other than image size?

Thanks,

Jim

[jkmorse]

Answered my own question.  Last night I took a large data set (72 900sec Ha's of M16) and ran the two side by side (gave me something to do while waiting for Hillary to speak ).  Virtually no difference and the drizzled image actually showed a touch more noise (the base image is extremely clean but with 18 hours of data I guess I should have expected a damned good SNR). 

For what it's worth,

Jim

[LarryC]

Jim,

I've been testing image sets of ~300 120sec ASI1600 mono images and find Drizzle does consistently reduce the noise in the final image.  I used to see a noticeable visual improvement when I ran slightly smaller sets of DLSR images, and well as noise reduction, but I see much less of a visual effect on my ASI1600 data.  In any case, the improvements are at least numerically sufficient for me to continue drizzling.

I have also been testing larger (smaller?) drop shrinks and now use 0.7 or 0.8 most often.

I have no idea what variables might be causes us to have different outcomes. 

Larry

[jkmorse]

Larry,

It may all come down to sampling.  I agree that drizzle is wonderful for my undersampled images, but for well sampled ones, I see no improvement.  What is the arcsec/pixel ratio on your set up? 

Jim

[LarryC]

Jim,

If the attachment worked, this is a screen shot of a highly zoomed in section of a 442 frame integration from my ASI1600 and RC6 at an image scale of 0.562.  Noise dropped from 1.177e-5 to 1.020e-5 (15%) and it appears to me that resolution has improved considerably in the drizzled version.  Interestingly, a recent integration with ~200 frames of another target at the same image scale and equipment did not show similar visual improvement.  That's the first time it hasn't been so apparent.  I can't explain the difference, but the attached image is very typical of what I see.

I've posted the same image over on CN in the drizzle thread in BII.

[jkmorse]

Larry,

Interesting!  Only problem is that you lose the ability to use the incredible Mure Denoise script but will retry.

Thanks,

Jim

[LarryC]

Jim,

It's clear that drizzle was developed to "fix" under-sampled images, but I don't see why that necessarily means it has no value to properly sampled, or in my case over-sampled images.  From CN, I think there is debate over what is meant be "resolution".  In my mind the example of the above clearly shows an improvement of resolution in terms of pixels on target - effectively the same thing as greater oversampling.  As a terrestrial photographer, resolution means to me a greater ability to enlarge without pixelation.  Some on CN argue that unless the FWHM is reduced by drizzling, that no increase in resolution has been achieved.  That makes no sense to me - that seems to me to be a measure of sharpness. 

Why can't one use the Mure Denoise script?  I haven't been using it, but I didn't know I drizziling would eliminate the possibility.

Larry

[vicent_peris]

Hi,

You can also do a drizzle x1. This way you preserve the original image resolution but you avoid any kind of artifacts introduced by interpolation. It's not only a matter of decreasing the FWHM values, you can also avoid any aliasing artifact that occurs when you have a slightly rotation (even if you're using the Lanczos interpolation).


Best regards,
Vicent.

[jkmorse]

Larry,

If you read Mike Schuster's documentation from Mure DN, he makes it clear that it simply doesn't work with drizzled data and I have confirmed this through my own testing.  And, since Mure is a superb denoising script, in my case I just don't think the trade off is worth it.  For images that do not use drizzle, Mure can be incredibly effective, cleaning both the background and the target areas incredibly well with no loss of detail even in faint nebulosity.  In my case, for my well sampled data, the trade off in using drizzle is that loss of Mure which is a trade I am unwilling to make.  If you haven't tried Mure, you owe it to yourself to check it out.  You will find it in Scripts under Noise Reduction.

Best,

Jim