Post-processing of compressed images with side information

This paper investigates the usage of side information for linear post-processing in subband image compression systems. In particular, side information is used to determine a custom synthesis filter bank for each image. It has been shown that perfect reconstruction filters are not optimal in the presence of quantization noise. However, most of the previous related work concentrated on optimizing analysis-synthesis pairs jointly, which requires very strong conditions. Limiting our attention to the decoder, we find optimal synthesis filters by solving a system of regression equations. Neither high-rate quantization, whiteness of the quantization noise, nor optimality of the quantizer is assumed. By allowing transmission of side information, we take a deterministic approach, thus finding the best reconstruction without any constraints such as stationarity. Optimal coefficients are independent of the internal machinery of the quantizers, and can therefore be used with “canned” software packages for post-processing purposes