Coding Prior-Driven JPEG Image Artifact Removal

Dongliang Cui, Youwei Pan, Wuzhen Shi, Yang Wen, Zicheng Liu, Yutao Liu

Abstract: Image priors play an important role in JPEG image artifact removal. However, most existing methods ignore the use of coding priors. This paper proposes a Coding Prior-driven JPEG Image Artifact Removal (dubbed CPIAR) method to improve the performance of JPEG image artifact removal. In the JPEG compression algorithm, because the algorithm divides the image into 8\(\times \)8 blocks and then performs quantization operations, this will cause serious blocking artifacts at the block boundaries. In order to make use of this information, we introduce a mask to represent the boundaries of image blocks. The introduction of this mask makes up for the lack of information about the JPEG compression process in the current deep blind method. We fuse this mask with image features, which can guide the model to focus the boundaries of image blocks, thereby better eliminating the blocking artifacts in JPEG images. In addition, we introduce the Degradation-Aware Dynamic Adjustment Block(DADA Block), which has better nonlinear expression capabilities and can dynamically adjusts the model based on estimated quality factors. Through this improvement we further enhancing its performance in handling JPEG images with varying quality factors.