1
\$\begingroup\$

This question is about image-processing methods to emulate the noise reduction and increase in details achieved by cameras with good built-in pixel shift (Sony A7 RIII, Pentax K-1 and others). I am assuming situations where pixel shift works well, particularly a good tripod, no wind, a static subject and enough time to take a few pictures.

There are several methods that I can think of, but none is really satisfactory:

  • Taking 4 identical images with a remote trigger and averaging. This effectively reduces the noise but doesn't bring up any additional detail in well exposed areas.
  • Adding minimal motion between the shots (e.g. touching the camera) and using superresolution. This eventually brings more detail - in addition to reducing noise - but it does require a lot more than 4 images and the process is really cumbersome.
  • Using a lens with 1.75x the focal length and doing a 2x2 panorama with 33% overlap. This requires just 4 images and the process is very straightforward. However, the depth of field shrinks and the aperture would need to be reduced by 3 stops to compensate - otherwise, focus stacking is required, making it as cumbersome as superresolution.

Are there any methods that would be effective and simple? Is there any software that would make superresolution a lot less cumbersome and more effective (e.g. something that would work directly on the RGB channels in the raw files and wouldn't need the intermediary expanded TIF images)?

\$\endgroup\$

2 Answers 2

1
\$\begingroup\$

The easiest way to get what you want is to use a camera with the feature built-in. Most tasks tend to be more cumbersome and time-consuming to do in post-processing than the same tasks in-camera.

If dpreview's description of Dynamic Pixel Shift is accurate, you should be able to take four hand-held shots, align, resize, and average them to produce equivalent results.

When using Dynamic Pixel Shift, image stabilization is switched on by default (a giveaway that it isn't doing single pixel shifts like its tripod-mounted counterpart) and the camera relies on the natural movement or shake of the photographer for four slightly different images. Stabilization remains on to ensure there is not too much of a framing difference between shots. The four files are then aligned and combined to produce one 'super resolution' image with increased detail, dynamic range and lower noise. Without getting into too much detail, sampling a scene multiple times with slight shifts to the image allows details to be localized with sub-pixel precision (since shifts are unlikely to be perfect multiples of one pixel).

I would use align_image_stack and ImageMagick. Here's how to create a super resolution photo with any camera describes using Photoshop. It also seems that using many more than four images is unnecessary.

Interestingly, the difference between the 4 image and 20 image super resolution examples is less noticeable. While there is some advantage to stacking more images, returns are diminishing in this case.

See A Look at Reducing Noise in Photographs Using Median Blending.

\$\endgroup\$
4
  • \$\begingroup\$ Thanks Xiota. Yes, I believe that the review of Dynamic Pixel Shift is accurate, and that it gives results similar to manual superresolution on 4 handheld images - definitely not as good as the real pixel shift. An yes, I use align_image_stack and convert but it is still a pain: very large file sizes, quite a few operations (I know it can be scripted but getting the align_image_stack options right can require some trial and error), pretty long processing times. In terms of image processing a 4 frames pano at longer focal lens is a lot lighter (when focus stacking is not needed). \$\endgroup\$
    – Come Raczy
    Oct 3, 2018 at 4:41
  • \$\begingroup\$ Median blending works well to remove moving elements but for noise reduction averaging is still better (ultimately, mean with clipping is probably the best). Beyond noise reduction, neither mean nor median seem to be able to recover detail lost by sampling in well exposed areas. \$\endgroup\$
    – Come Raczy
    Oct 3, 2018 at 4:45
  • \$\begingroup\$ I made a few attempts a while ago, just to try, and decided it wasn't worth the effort because of what you describe – large file sizes and long processing times. Also, most of what I photograph moves, even when I think it doesn't. \$\endgroup\$
    – xiota
    Oct 3, 2018 at 5:53
  • \$\begingroup\$ You can try creating the intermediate files on a RAM disk. How well that would work depends on how much spare RAM you have. \$\endgroup\$
    – xiota
    Oct 3, 2018 at 5:55
0
\$\begingroup\$

Are there effective and simple methods to emulate pixel shift?

  1. Fake it with an online (or APP) image enlarger that can increase your size as much as 10x (or more), very fast with reasonable results. Obviously nowhere near as good as using more pixels to begin with, but it works with one photo (which might be all you have).

  2. Take over a dozen hand-held shots holding the camera relatively still and PhotoShop the images to increase the resolution:

    1. Import all photos as stack of layers
    2. Resize image to 4x resolution (200% width/height)
    3. Auto-align layers
    4. Average layers
  3. AutoPano Video will accept a video input, or just hold the shutter button and make a video from high resolution stills. You can hold the camera relatively still or move it slowly, then feed the video in and use the software to crop an image from the extents of the input.

For best results use hardware such as the "R1 Adjustable Tilt Ring Mount & Rotator Mini Package (F6161V)" or the "NN6 w/Nadir Adapter D10 (F6002)" to take images without parallax and feed the best quality images into the above software, that reduces the warp and enhances the final result.

Nodal Ninja

\$\endgroup\$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.