Recover image coding loss with LMS filtering

Since the coding residues correlates with the coded image, we propose a model for compression residue estimation with the coded image. When used for postprocessing, it results in a LMS (least mean square) process which can be implemented as filtration similar to most existing techniques, but gives more theoretical insight and restores residues without any assumption of the coding distortion (i.e., with higher level in generality). Determination of the critical bit rate is not required since it works for all bitrates, and is applicable to any image, codec and artifacts. The approach avoids the difficulties in image region discrimination and filter switching that are encountered otherwise. The proposed model results in around 0.5 dB for average PSNR improvement consistently over the exhaustive range of bitrates (0.15~4.0 bpp), with more gain at a lower bitrate, for various images.