Deep learning-based compressed image artifacts reduction based on multi-scale image fusion

One of the visually noticeable compression artifacts in block-based image/video compression platforms is called blocking artifact. Several post-processing methods were presented to reduce such kind of artifacts. However, most methods in the literature often induce visibly blurring artifacts. The paper presents a deep network to eliminate image compression artifacts (usually denoted by image deblocking) based on image fusion in multi-scale manner. Recent deep learning-based related methods usually learn deep models using a loss function in per-pixel manner based on explicit image priors in order to directly produce clean images. In place of existing deep learning-guided approaches, the problem is reformulated in this paper to the learning of the residuals (or artifacts) between the received images and their corresponding clean images (ground truths). In the presented deep framework, an input image is first down-sampled while naturally reducing the blocking artifacts. Then, our multi-scale image fusion model is used for fusing the different down-scaled versions (of less artifacts) with the input image (with severer artifacts) to estimate the blocking artifacts. Then, by deducting the estimated artifacts from the input image, the blocking artifacts can be significantly eliminated and most original image details are preserved simultaneously. The presented method is well applicable to any vision-based computer systems with digital visual codec embedded.