>DCT implementations are usually not bijective even without quantization
imho their precision is limited to ~12bit for 8 bit data, (ijg do so) if i remember right
What precision need to make it losless(bijective), i thought 16bit int is enough but now i'm doubt about that.
>To get some estimation you can compare eg. Shkarin's BMF with png, or >PPMD vs LZMA on raw image data.
tested on some drawing scans
Code:
size c.time
1953439 5 6-o2.7z
1955696 4 6-o3.7z
1960722 5 6-o4.7z
1985729 7 6-o8.7z
2218561 144 6-lz.7z
56431320 6.tif
4639565 6 4-o3.7z
4643738 8 4-o4.7z
4676973 4 4-o2.7z
4716066 15 4-o8.7z
5328316 178 4-lz.7z
56433616 4.tif