[daala] New entropy coding - faster than Huffman, compression like arithmetic

Jarosław Duda dudaj at purdue.edu
Sun Dec 29 10:04:06 PST 2013


There is a new approach to entropy coding (no patents) - the current 
implementation has 50% faster decoding than Huffman (single table use 
per symbol from a large alphabet), giving rates like arithmetic:

The cost is that we need to store coding tables for a given probability 
distribution: let say 1-16kB for 256 size alphabet.
So we could use separate tables for different contexts, as long as their 
number is relatively small - probably could be used to speed up a few 
times compression of DCT coefficients, maybe also prediction?

Jarek Duda

dr Jarosław Duda
Center for Science of Information, Purdue University, USA

More information about the daala mailing list