Post-processing a block-based discrete cosine transform (BDCT) encoded image requires solving two seemingly contradictory tasks of suppressing discontinuities at block boundaries and enhancing edge and texture
details. This paper approaches this problem by combining the existing algorithms which are specialized to each individual task. The re-application of JPEG applies BDCT coding to pixel-wise shifted versions of the input
BDCT encoded image and takes the average of these re-encoded images after shifting them back to the original positions. This step effectively restores continuities in block boundaries while tending to keep the existing details.
Then, a set of regressors are trained based on example pairs of these artifact-suppressed images and the corresponding clean images such that missing high-frequency details lost during the DCT are restored. Furthermore, two generic image models which take into account the leptokurtic nature of natural images in the wavelet and spatial domains, respectively are adopted so that block artifacts remaining after the regression step are removed and the major edges are enhanced. Comparison with the existing post-processing methods shows the effectiveness of the proposed method.