MPRNet
MPRNet is a multi-stage progressive image restoration architecture introduced at CVPR 2021 that achieves state-of-the-art results for image deblurring, deraining, and denoising. The model addresses the challenge of balancing spatial details with high-level contextual information by breaking the recovery process into manageable steps across multiple stages. It utilizes an encoder-decoder structure to learn contextualized features while maintaining a high-resolution branch to preserve local details. A key innovation is the per-pixel adaptive design using supervised attention modules to reweight features at each stage. The architecture ensures robust performance through comprehensive information exchange, featuring both sequential progression from early to late stages and lateral connections between processing blocks to prevent information loss. Validated on ten diverse datasets including Rain100L, GoPro, and DND, MPRNet demonstrates significant PSNR improvements over existing methods. The implementation is buil