Raphael Canut
2017-05-25 18:28:17 UTC
Hello,
I spoke a little quickly when I proposed a £1000 rewards... because there
is still some "easy" processing to increase compression in the NHW codec,
like increasing wavelet coefficients threshold, decreasing again residual
coding.I will try to code it tomorrow and this weekend, and will let you
know what it gives.
A new idea I'm thinking of not to have too much aliasing and annoying
artifacts, is if a wavelet coefficient between let's say
6<wavelet_coeff<thereshold, has a corresponding wavelet coeff in the
next-order decomposition bigger than threshold2 then this wavelet
coefficient is not zeroed...
I will try to do this "basic" work in the next days and will let you know
if it is acceptable and competitive.I think I could add ideally with this
another -5,-10,-15Ko in compression but for higher compression I think I
must introduce a new approach like GANs maybe (generative adversarial
networks) but I don't know if it will fit my speed and low-memory
consumption goals...
Stay tuned.
Cheers,
Raphael
I spoke a little quickly when I proposed a £1000 rewards... because there
is still some "easy" processing to increase compression in the NHW codec,
like increasing wavelet coefficients threshold, decreasing again residual
coding.I will try to code it tomorrow and this weekend, and will let you
know what it gives.
A new idea I'm thinking of not to have too much aliasing and annoying
artifacts, is if a wavelet coefficient between let's say
6<wavelet_coeff<thereshold, has a corresponding wavelet coeff in the
next-order decomposition bigger than threshold2 then this wavelet
coefficient is not zeroed...
I will try to do this "basic" work in the next days and will let you know
if it is acceptable and competitive.I think I could add ideally with this
another -5,-10,-15Ko in compression but for higher compression I think I
must introduce a new approach like GANs maybe (generative adversarial
networks) but I don't know if it will fit my speed and low-memory
consumption goals...
Stay tuned.
Cheers,
Raphael