I read the paper. I really wonder how it is possible, that it works with lines treated independently...
Moreover, I am little unsure in the part 3.5 of the paper...maybe my view is then wrong and needs correction:
Is your goal JPEG or MPEG or both?
In case of MPEG I think it could be interesting to incorporate weighting by quantizers (two adjacent blocks with different quantizer can differently contribute to the discrepancy). Interior discrepancy values will also depend on quantizer (higher q=smaller interior discrepancy, bigger boundary discrepancy).
Also, error magnitude can be estimated using quantizer and matrix values.
|