Mipmap stacking

[vlaiv]

Here is another interesting idea that I came up recently.

Haven't implemented or tested it yet, just wanted to share the idea and discuss it a bit.

Mipmaps in computer graphics are pre calculated progressively smaller images that are easier to handle / take up less space and can be applied to object if it is smaller - as they contain less detail. This is usually done in 3d graphics with large scenes where object can be far and thus small and close and thus large - think for example of 3d space simulator and you approaching space station. At first, space station is far away and you don't see / need much detail to be drawn. As you approach - you need more detailed view of it and you use higher quality texture.

What does this have to do with stacking?

Well, it has to do with binning and resolved detail. Imagine you have following scenario:

You have SNR map and your bin x1 stack. SNR map tells you SNR for each pixel in the image. You proceed to create x2 binned version of the image, x3 binned version of the image and x4 binned version of the image. You then resize all of those to original size by some resizing algorithm (in reality - this binning and resizing will be done at stacking stage by using sift bin and resampling to match original size).

In the end - you combine final image like this: You set thresholds at some points - say >10, 5-10, 2-5, <2.

For each pixel with SNR>10 you keep original pixel value - of bin 1 image

For each pixel with SNR in 5-10 range you replace it with corresponding bin x2 image pixel

Similarly you do for 2-5 and <2 pixel values replacing them with bin x3 and bin4 images.

This will create very smooth looking image in background and very detailed image where you have enough signal. It is like denoising step performed during stacking stage.