Probabilistic Compression Artifacts Reduction Using Self-Similarity Based Noise Region Estimation

During compression artifact reduction process, original information as well as noise has been commonly removed, and this side effect should be importantly considered. In this paper, we propose a novel post-processing approach to alleviate the side effect of noise reduction while still reducing compression artifacts successfully. After compression artifact removal using conventional methods, we examine whether the denoised region is actually noisy or not, exploiting the relationship between noisy image and artifact reduced image. Then, the probability of a pixel to be noisy is calculated based on the noise region estimation, and a final denoised pixel is obtained by a weighted average between noisy and denoised signals with the probability. Experimental results show that the proposed method is more effective in preserving texture region while still reducing the compression noise.