I cannot see how this should work... Some improvement? Yes. 9x improvement? No. There were other approaches in the past (I do not recall their names) with 4x4 arrays, etc. which promised less and ultimately failed (at least to my knowledge).
You are thinking of Pelican Imaging, 9X Wow quite the claim. Might look a bit weird having 1/9th the resolution in the monochrome sensors morphological shadow regions of the image though
To my very best understanding (up to now) this an approach to do sub-pixel interpolation by means of neuronal networks. This means the algorithm was trained for a bigger set of patterns and structures from real world images and when used will deliver the most likely patch for blowing up the resolution in a sense making, likely fashion. As with many approaches in this area of image upscaling the driving idea is likeliness - and by the common understanding of the used neural network a very high likeliness for the result should be achievable. It is not meant to be an exact sensing but it definitely is rather smart way of doing a guess based upon "experience" from comparable previous data.
Without an easy to access option for very own testing (web interface? exe/lib/SDK for download) i assume the company simply falls too short for the masses or even the academics world. maybe they have something like this to offer for their supposed customers - might that be security equipment manufacturing companies.
I cannot see how this should work... Some improvement? Yes. 9x improvement? No. There were other approaches in the past (I do not recall their names) with 4x4 arrays, etc. which promised less and ultimately failed (at least to my knowledge).
ReplyDeleteYou are thinking of Pelican Imaging, 9X Wow quite the claim. Might look a bit weird having 1/9th the resolution in the monochrome sensors morphological shadow regions of the image though
DeleteTo my very best understanding (up to now) this an approach to do sub-pixel interpolation by means of neuronal networks. This means the algorithm was trained for a bigger set of patterns and structures from real world images and when used will deliver the most likely patch for blowing up the resolution in a sense making, likely fashion. As with many approaches in this area of image upscaling the driving idea is likeliness - and by the common understanding of the used neural network a very high likeliness for the result should be achievable. It is not meant to be an exact sensing but it definitely is rather smart way of doing a guess based upon "experience" from comparable previous data.
ReplyDeleteWithout an easy to access option for very own testing (web interface? exe/lib/SDK for download) i assume the company simply falls too short for the masses or even the academics world. maybe they have something like this to offer for their supposed customers - might that be security equipment manufacturing companies.