Noise: A Nerd's Approach

A few weeks ago there was a thread here:

https://photography-on-the.net …showthread.php?​p=17356280

entitled "How-to: Making an underexposed low-ISO Canon image look like an underexposed low-ISO Sony image."

I gather the intent of the post was to demonstrate a workflow that could be used to eliminate the various kinds of noise for which the Canon sensor is known with some processing techniques, including FFT filtering for the banding noise, etc.

Here is another approach, although I am not aiming to make the Canon image look like a Sony, or any other image, just throwing the idea of trying to increase signal-to-noise ratio and fight the pattern noise gremlins.

Right up front I am going to state that this is a demonstration of a long-used technique in astrophotography and clearly may not be useful in most everyday shooting. However, I am a nerd and it's all about getting your nerd on - so here goes.

This is the scene/final image I set up and processed:

I apologize in advance to the astrophotographers on the forum who will read this and cringe at the oversimplification and likely mis-worded and glib explanation of the workflow. I mean no harm.

I used the astrophotographic workflow for calibration and image integration. This process is predicated upon sampling the camera's noise characteristics and optical characteristics and creating files that are subtracted from the scene image files to remove things like pattern noise and vignetting, etc. The terms that describe the data/image types are:

dark
bias
flat (field)
light

Dark frames and bias frames are used to characterize the sensor and electronics noise. Flat field frames are used to characterize light falloff and other optical flaws (like the LCC images used in CaptureOne) and the lights frames are the actual images of your scene.

For each of these four image types you shoot multiple, identical images (10-20 per image type). The general idea is that averaging or combining the 10-20 images into a "master" image will permit the master dark, master bias and master flat to be removed from each of the light images - this is all done before debayering so that each pixel location on the sensor is accounted for in the CFA pattern, and no interpolation has taken place. This is known as "calibration."

Once each light image has been calibrated, the calibrated light images are integrated (combined) into a single master light image. This is the image that you can then use as you would normally to edit and produce a final image, with more signal and less noise. The steps required to produce a true astrophotographic image go well beyond this preprocessing, but for my purposes in shooting this scene, I was able to use the master light image as I would a raw image.

I did all of this fancy stuff in PixInsight. I exported the resulting master light image as a 32bit TIFF and then opened it in Photoshop. I used the Camera Raw filter to tone the 32bit image, pushing it about 3+ stops and boosting the shadows significantly. I was not too concerned with aesthetics and used no noise reduction. I used all DEFAULT settings in PixInsight for this preprocessing, which likely yields less than ideal results. Even with those settings, I was able to achieve significant results.

I also opened one of the original raw files (one of the lights image raw files) and applied similar settings in ACR and converted into a 16bit TIFF in PS. I color matched this output to my master lights image to make the images visually consistent.

What follows are 100% crops comparing three image states:

1) the zeroed raw image
2) the ACR conversion of one of the raws
3) the calibrated and integrated image

kirk

Highlight area (styrofoam head right next to the lightbulb):

Head in deep shadow:

Here's one more example:

Nice one Kirk.

I've often wondered if those astro techniques could be used for other types of photography. However, because it's really only useful on totally static subjects (and that's something I never shoot) I've never been tempted to try it. Good to see that it works so well.

Excellent noise control.

Another of your comprehensive and detailed tutorials that you obviously put a lot of time and work into. The POTN community appreciates it.

Thanks for the comments! While this was an experiment to demonstrate the concept, I think it might be a worthwhile technique in some circumstances. My limited understanding of this process in astrophotography indicates that the acquisition of dark frames, especially, is required on a per-shoot basis in most cases because exposure times are long and camera sensor temperature plays an important role. The assumption I would test for cases like the one I shot here is that exposure time is so short compared to astrophotographic times, that the acquired dark, bias and flat frames can be applied generically over a wide range of conditions. That is, you could have a library of master darks, bias and flats for only a few different shooting conditions (or lenses) and simply shoot 10-20 lights images of your scene and process them with the appropriate generic master dark, bias and flat.

Sounds like another experiment! Will post here.

Kirk

Could you generate the first three sets of filter images using the camera and lens combinations that you are likely to be using in the "real world" separately from the actual wanted image. Could you then apply the filtering to an image created from a single exposure? Or do you actually need to have multiple exposures of the required image, in order to average out a lot of the noise? If you could apply the corrections to a single exposure it could actually make this sort of approach far more useful in real world applications with moving subjects.

Alan

BigAl007 wrote in post #17381696
Or do you actually need to have multiple exposures of the required image, in order to average out a lot of the noise?

Using dark frames is a great way of getting rid of systematic noise (that's always in the same place). But it won't do anything for random noise. Stacking multiple images will increase the signal:noise ratio (by the square root of the number of stacked images - so shooting 25 'lights' will reduce random noise by 5x).