Physics-Guided Image Restoration

The field studies how to recover useful images when the imaging process has introduced noise, blur, missing measurements, or other distortions, across settings such as medical imaging, microscopy, remote sensing, and low-light photography. Methods combine explicit models of image formation—such as optical point-spread functions, atmospheric scattering, and sensor noise—with neural networks, diffusion models, unrolled optimization, or self-supervised learning. Synthetic physics-based training data and scan-specific or test-time adaptation are common ways to address scarce paired examples and changing instruments or conditions. A central challenge is ensuring that physical guidance is genuinely valid and improves real-world generalization, rather than merely acting as a useful learned bias; evaluation increasingly considers uncertainty, physical consistency, and downstream tasks alongside visual quality.