Deconvolution-Just can't Seem to get good results

[sreilly]

I have uploaded an image of M82 that consists of 23-20 minute luminance images taken with my 12.5" RC @native f/9 with my STL-11000M camera. All of the images are =<2.5" seeing as measured by CCDInspector. The images were calibrated, aligned, and combined (average, Windsorized Sigma Clipping, Clip Low/High Pixel/Range). The uploaded image has had nothing further done. Normally I would have cropped the uneven edges, the images are dithered, and performed DBE (usually twice using both Subtract and Division methods). As I see from all the tutorials I've seen on deconvolution, the image should be linear to use deconvolution.

I'd appreciate anyone who cares to try to use deconvolution on the image and explain the methods they used, screen shots would help explain a lot, have a go and post back with their results. I have tried using the Deconvolution Preview Script but obviously don't fully understand what the starting parameters should be or how to go about getting these values. I would think using the Dynamic PSF function would get me in the ball park. The direct link for download, 33.16 MB, is http://www.astral-imaging.com/M82-23-L.zip

Anyone able to shed some light on this for me would be a great help. Maybe I'm a bit too slow for this method!

[dayers]

What results are you trying to achieve with this data with deconvolution? From what I've read, most people are trying to "unsmear" stars messed up by poor seeing.  I downloaded your image and took a look with DynamicPSF. The stars look pretty darned good to me. I think they are undersampled a bit, and that might make it difficult to tighten them up. I don't know.

The nebulosity of the galaxy is something else, and I just don't know what to expect there.

Thanks for sharing your data. I'll give it a go.

Dave

[sreilly]

The resolution is .644 arc seconds per pixel at 1x1 bin for the 12.5" RC 2f/9 and the STL-11000M.

[pfile]

after on the left, before on the right. is this what you are after? or have you already achieved something like this? i think it got just a bit crunchy but i did not experiment a lot with different settings.

[Juan Conejero]

Hi Steve and all,

First of all, I'm sorry for not being active on the forum lately. I've been ill for the last two weeks. Nothing serious; just a bronchitis but I can't remember ever being so ill. Now I'm starting to feel better, so I'm afraid you won't get rid of me so easily 

I've made a brief tutorial to describe some easy steps with this image. The first one is to find a good estimate of the PSF with the DynamicPSF tool:

Click for full size image

Note that I have sampled about 50 stars (less would suffice) around the object of interest. Since the image shows varying star shapes throughout the field, I have sampled the region close to the main subject for deconvolution.

The PSF model image has been generated by selecting all PSF measurements and clicking the export synthetic PSF button. This is definitely the best way to generate a robust PSF model with the DynamicPSF tool. I strongly discourage using average PSF parameters for this purpose.

The next step is deconvolution. I have used nearly default parameters with the Deconvolution tool. 20 iterations of the regularized Richardson-Lucy algorithm, with default regularization parameters and no mask. The only slightly critical parameter has been the global deringing threshold. This is the before image shown with custom screen stretch parameters (for better inspection of the galaxy core):

Click for full size image

and this is the image after deconvolution:

Click for full size image

Note that I have used no mask in this case. Wavelet regularization and global deringing have been sufficient to prevent the worst side effects of deconvolution for low-SNR areas (noise intensification) and small-scale bright features (ringing around stars and other bright structures).

The third step is a nonlinear stretch with HistogramTransformation. I have applied automatic STF histogram parameters.

Click for full size image

Next, I have applied HDRMultiscaleTransform with 5 wavelet layers:

Click for full size image

A side effect of the HDR multicale transform algorithms is that bright structures that are close to saturation in the original image can cause artifacts. This can be seen as 'concave' bright star cores in the screenshot above. To fix this problem we need a simple star mask:

Click for full size image

This mask is very easy to build because we only need protection for the brightest stars. This is the mask active and shown for the image:

Click for full size image

and here is the result of HDRMultiscaleTransform applied with the star mask enabled:

Click for full size image

Finally, some noise reduction may definitely help this image. Let's keep it classical with the good old ACDNR tool. Before:

Click for full size image

and after ACDNR:

Click for full size image

Hope this helps.

[FunTomas]

Juan, simply fantastic example!!

[ManuelJ]

I still don't like DPSF results

[sreilly]

Thanks Juan and I sincerely hope you are feeling better. I know it takes a bit of effort to do these especially if you're not feeling up to par. Sometimes it takes this kind of explanation for me to get the process. I mimicked the process with your settings and, of course, got the same results. What I need to do now is play with the settings, such as HDRW and see what the differences are. As much as I love PI, what intimidates me most is the extreme control some of these functions have and trying to understand what the settings do to the image. Generating the star mask is a fine example. I've found that using a scale of 7 seems to work best for me but I had never used the Binarize choice or had the Threshold set as high before. Unchecking the Growth choice is also something I've never done.

Inverting the mask has been how I've learned to apply HDRW so that's not new to me. Is there an advantage of using ACDNR over the MMT noise reduction process? I realize that MMT is done on linear data so it would have to be before using HST. I was also wondering when in this process you would apply DBE?

Again, I really appreciate the time and effort you put into these projects.

Regards,

[swag72]

Thanks Juan, that is a really good tutorial and one I shall be tagging for future reference.

[mmirot]

Why didn't you mask those bright stars before deconvolution?

Max