Would it be worthwhile to compress the edges separately? I.e., use edge-detection + dilation + quantize + mask to get the edges, then compress that less aggressively, and overlay over the aggressively-compressed image?