I often see people stacking small numbers of photos in an attempt to improve image quality. Sometimes this works as expected, and other times…. not so much. One of the major issues with this technique is that it can actually yield lower image quality when the individual frames are subject to atmospheric distortions or occasional vibrations in the camera and mount system itself.
[ Left to right: One of the less blurry stills, 100 image stacked, one of the more blurry stills. ]
For example, here are 300 shots ( 100% crops ) of a very close subject through a 300mm ( 450mm equiv ) lens. Its focused very close to the minimum focus distance, and wide open ( hence the various image flaws. ) But what we are focusing on is the undulating “mirage” like appearance of the image.
Video: 100% Crops, As Shot
Here is the same source stills, stacked 10 images at a time.
So what is happening? The initial video, while certainly an undulating mess, does seem to have portions / frames which seem sharper. Sometimes the sharp portions are very locallized. The stacked images, while eliminating some of the visible distortion, are also more uniformly blurred, and still some of the frames are far more blurry than individual samples from the stacks. For example:
A good individual frame, hand selected:
Compare to one of the more blurry frames, hand selected:
And here is a 100% crop of 100 images, stacked:
So even with stacking 100 images, we don’t really improve over one of the subjectively better original stills images. By averaging all the images, you also bake in both the best and the worst of all the images.
So what to do? If there were only a way to select and stack the better frames and reject the blurry ones. Well of course there is. But even better would be to chop the image up into small pieces and reconstruct our stacked image from the best of the best for each of these small pieces.
None of this is new, or original, and its pretty standard for astrophotography – but these techniques could be used to improve telephoto shots destroyed by atmospheric distortions.
Approach 1 – Identify and stack the least blurry frames.
Obviously if you had an algorithmic measure for how blurry an individual still is, you could rank all of them, and then then stack, say, the top 10%.
Approach 2 – Chunk the image into little regions, and for each identify the best 10% and reconstruct the image from them like a patchwork quilt. This handle the situation where an individual frame has both very good and very bad regions which would be deselected by Approach 1.
Approach 3 – Retroactive Adaptive Optics – Sort Of…
One of the major issues with the first two approaches is that it does not handle instances where regions are sharp, but suffer geometric distortions, such as those seen in first video. The sum of these sharp images would still be very blurry when stacked.
The solution is to take the average of the “good” frames ( or regions ) from a blurring standpoint, and then compare the sharp individual frame to distort the individual still to match the baseline. The corrected image would then be stacked forming a distortion corrected, and less blurry, final image.