![]() Once interpolated, the debayer algo (AHD) leaves more structured noise which is much more difficult to eradicate. One thing I did find was that the denoise worked much better on pure pixel noise - i.e., RAW. The relevant files are all there on the drive if you want to look deeper. NOTE: training the model takes a long time, but once you have it, denoising an image takes only 2-3 secs.Īs above, but comparing the 300 epoch results with low-ISO version. For the 6d - which is relatively low noise anyway - there was a small improvement over the 30 epoch quicky version, but I was gobsmacked how well it worked on the very noisy G9 image. Went out climbing this morning, so left the computer training a 300 epoch model (Canon). The compares the native model result with a low-ISO (ISO1600) version of the same image (a sort of ground truth). It looks like the result are noticeably better with the native model. I initially ran the training for a minimal 30 epochs (~30 mins) on the 6d and the G9 images and then ran the images through their 'native' and 'foreign' network models. I've put the results on.Ĭompare-h.jpg (100%), compare-h2.jpg (400%) So I generated some noisy ISO25600 images on both my full-frame Canon 6d and my m4/3 Lumix G9, battled with installing all the right versions of ML libraries (another story) and did some testing. I found a derivative algo called Noise2Void which is simpler to train and also is available as a plugin to ImageJ. The code is available, but a bit awkward to use. I believe Topaz denoise is based on an ML algorithm called Noise2Noise. My interest was peaked by this tread, so being too stingy to invest in Topaz, I had a dig into the similar open source codes. That doesn't work so well if you have foreground though - the foreground isn't affected as much by airglow. Of course for most DSO objects you'll subtract out the airglow in the workflow, so it won't be obvious. Show a green sky to most people and they'll say "Nah, that's wrong!" But that's the true colour of airglow. * The most obvious example is if you take colour balanced image of a proper dark sky, most people expect it to be dark blue (in so far as it has any colour at all). ![]() If you are using any filters like UHC or pollution filters, or a modded camera, then all bets are off - choose whatever white balance you like, they won't be 'true' colours in any normal sense of the word. Target a suitable star field, look up the expected colours in say Stellarium and then adjust your red/blue multipliers until they are as close as you want. ![]() If you want accurate, then assuming you are not using any filters, I tend to think the best way is to use star colours to calibrate your lens/sensor. In reality most amateur astrophotographs are more art than science. OK, first you have to decide: Do you want the colours to be accurate or do you want it to look nice/interesting? They are not always the same*. ![]() If you are photographing a landscape, not DSO, how do you choose WB? ![]()
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |