@cameroncamera: I see what you’re going for here, I think. That would be a process for upscaling a digital image, but if I’m reading it right, wouldn’t there be an issue with duplicating image detail across pixels, causing another form of interpolation smearing? Each image is expanded so that there is a one pixel gap between each pixel. If A is a pixel and B is an empty pixel, the result would be this:
ABABAB
BBBBBB
ABABAB
BBBBBBThe second frame would then be shifted one pixel to the right, filling in the B spaces in the 1st and 3rd rows. The third frame would be shifted down, so that half of the 2nd and 4th rows would be filled, and the fourth frame would be shifted down and to the right, completing the picture.
So far that’s your process, as I understand it. Now imagine that the image showed a red light in the upper left corner of the image, taking up only one ‘A’ pixel. If each of the four frames showed relatively similar detail, the upscaled image would show that single pixel of red repeated four times in a box configuration.
I don’t have any idea about upscaling a digital image, since the pixels are the detail. Perhaps the only way to really upscale digital content like that would be through an adaptive learning algorithm such as the ones being developed by Google, wherein it identifies common objects and upscales their detail with images culled from a library of images. http://newatlas.com/google-raisr-image-upscaling-super-resolution/46434/
Ok yes I am glad my description made sense. Mine is an untested concept, though it may me exactly the manner in which typical temporal super-resolution algorithms work. I doubt that I am coming up with anything new. I think, however, that film-based image sequences could result in better upscaling than images originating from a digital sensor. (I do understand that a digitized film frame sequence is captured with a digital sensor, but I have hunch that the random film grain and slight gate weave of analog image capture would work advantageously for temporal super-resolution.)
You do bring up an excellent point about tiny image details that are effectively 1 pixel in size being inappropriately quadrupled when upscaled. In my example I presented a strict sequential approach, where the first pixel of the first of four frames is always placed in the upper left position on the larger canvas, then the second frame slots into the next available position, and so on. But, what if the super-resolution algorithm could intelligently place the expanded frames to maximize detail? Say, the super-resolution algorithm could talk to a stabilization algorithm? Perhaps analysis of random film grain and slight gate weave amongst the selected film sequence frames could rearrange the upscaled pixels to maximize real detail instead of magnifying errors in image detail.