Compressed Image Artifacts Removal with U-Net and Attention Mechanism Net

Traditional block-based image compression introduces block artifacts. The evolution of deep learning has simplified the process of image artifact removal. However, the diverse quality of compressed images post-compression means that a universal model cannot uniformly enhance image quality. Furthermore, the lack of attention mechanisms in the estimation of image quality leads to inaccuracies. To overcome these challenges, we introduce the U-Net and Attention Mechanism Network (UAMN) for Compressed Image Artifacts Removal. By training a single U-net, we enable blind artifact removal from images. We also incorporate attention mechanisms into our Quantization Parameter (QP) estimation to enhance image quality. Our testing shows a 11.17% performance improvement over traditional JPEG images, effectively upgrading the quality of compressed images.