jacobrosenthal
commented
7 years ago @tradica published some early thoughts http://tradica.com/memory.io/
"When images (spacial domain) are transformed into component frequencies, we can know what type of blur is dominant. Answers questions like, did the subject move, did the laptop move, did the phot happen while the laptop was opening ... three general types, linear, radial and zoom Nothing's exact and it's likely a combination the reduction process is a virtual committee and a stack of binary tests conclude in the selection/results five to ten tests should determine good images Quantifying result images of transforms and certain pixel zones of those images can reliably determine if it's blurry Blurry images can be useful/ Knowing the lens profile can isolate foreground subject from background. If a similar enough image is captured in same environment, the blur data can be applied to the clear image since there's a bit of motion in the blur image, distances can be modeled. It can be a bit fuzzy, but once a few tricks are piped together there's some awesome power just added two links to bottom of that page""
Thinking about trying to quantify image quality in memory io shots Like blur/focus/light and multiple faces 'interesting' backgrounds, ie not single color, these represent a possible memory of a location
After some searching a term I found is "nr iqa" (no reference image quality assessment) or blind quality assessment Though most of that research is around finding suitable confession algorithms and % that don't distort too much.
thinking you give each metric a score maybe? more interesting can float to the top then
frankly though nueral networks are the thing for this nowadays