So yesterday I had a quick look on DLI image format. It turns out to be somewhat related to video codecs (and JPEG of course): there’s 8×8 fast integer DCT approximation, quantisation and bit coding of the block. And bit coding is the most interesting part reallyâ€”this format employs binary model with old-school arithmetic coding and context selection for the model; coefficients are coded as first an array of coded coefficients flags (plus a flag for last coded coefficient), each non-zero coefficient has an additional flag to signal whether it’s larger than one and in that case it’s coded as unary code (bits coded with arithmetic coder of course).
And I still don’t like the notion of “let’s make our video codec I-frames into an image codec” (the reverse of “motion
<image format>” is not much better but at least it makes sense for intermediate formats). Images and video codecs have different use cases and required features but I think I ranted about it once.