Posted By
Degauss
on 2018-11-01
12:35:32
|  Re: DFLIConv2
you're right: its hard to believe CLCL is always the best way to go 
I didn't dig as deep as you did but my attempts with "adaptive" didn't show any benefit over the straight CLCL-config - at least for pictures i tested. I figured in the end i was only trading error in one place for the same amount of error in other places and eventually i lost interest in pushing this further (i started with random changes + threshold accepting. quick results, but also very quickly i reached a dead-end here).
Still, i had some ideas, but neither the time nor the brains to give them a try. I bet you came across these ideas already yourself, but i got the feel i need to dump them here anyway:
1) i always used sum of absolute squared error (compared to an "ideal" dfli-picture) for measuring fidelity. i believe there is a lot of potential in other measurements. Dunno, something more perceptive or some model that factors things like pal-color-bleeding in...
2) something probably not that difficult would be to add weighting to the errors to help the optimizer take the right decisions. that could be a user-provided region-of-interest or something based on image-features (e.g. sharp corners)
|
