DPIR (Denoising Prior-based Image Restoration)

DPIR (Denoising Prior-based Image Restoration) represents a paradigm shift in low-level computer vision by leveraging a Deep Plug-and-Play (PnP) framework. Developed primarily as a research-grade tool, DPIR integrates a Half-Quadratic Splitting (HQS) algorithm that decouples the image restoration process into two distinct sub-problems: a data sub-problem (likelihood) and a prior sub-problem (denoising). This modular architecture allows a single pre-trained deep denoiser (typically the DRUNet backbone) to solve various inverse problems including SISR (Single Image Super-Resolution), deblurring, and colorization without retraining for each specific task. By 2026, DPIR remains a foundational modular component in high-end medical imaging and satellite post-processing pipelines due to its mathematical interpretability compared to end-to-end black-box models. Its ability to handle spatially variant noise through noise-level map inputs makes it particularly robust for real-world scenarios where camera sensors produce non-uniform artifacts. As an open-source framework, it has been widely adopted by the scientific community to provide a benchmark for PSNR and SSIM metrics in restoration tasks.