Content-adaptive postfiltering for very low bit rate video

We propose a postfiltering algorithm which adapts to global image quality as well as (optionally) to semantic image content extracted from the video sequence. This approach is in contrast to traditional postfiltering techniques which attempt to remove coding artifacts based on local signal characteristics only. Our postfilter is ideally suited to head-and-shoulders video coded at very low bit rates (less than 25.6 kbps), where coding artifacts are fairly strong and difficult to distinguish from fine image detail. Results are shown comparing head-and-shoulder sequences encoded at 16 kbps with an H.263-based codec to images postfiltered using the content-adaptive postfilter proposed. The postfilter manages to remove most of the mosquito artifacts introduced by the low-bit-rate coder while preserving a good rendition of facial detail.