Instead of searching for quality using bisection method, which is slow, I'm trying to find exact quality needed by modelling jpeg quality vs dssim ratio as a curve, and finding the desired point on that curve.
I think I've told my math teacher that polynomials are stupid and I'm never going to use them… but…
The curve is a polynomial constructed from 3 points, and the points represent jpeg-quality to dssim relationship from the last 3 runs (default points are used for the first couple of runs).
It needs a bit of tweaking, but when it works it finds exact quality in 3-4 steps.
https://github.com/pornel/imgmin/compare/pornel:0fe0e25...pornel:d0b9172
Instead of searching for quality using bisection method, which is slow, I'm trying to find exact quality needed by modelling jpeg quality vs dssim ratio as a curve, and finding the desired point on that curve.
I think I've told my math teacher that polynomials are stupid and I'm never going to use them… but…
The curve is a polynomial constructed from 3 points, and the points represent jpeg-quality to dssim relationship from the last 3 runs (default points are used for the first couple of runs).
It needs a bit of tweaking, but when it works it finds exact quality in 3-4 steps.