augraphy

Why we included this project

Paired clean-and-noisy document images are scarce, and that scarcity is the problem Augraphy was built around. It takes a clean document image and runs it through a configurable pipeline that reproduces the wear real office equipment leaves on paper, including dirty laser or inkjet printers, low-resolution fax machines, bleed-through, folds, shadows, and scanner noise. Every degraded output comes with a perfect ground-truth copy because the pipeline starts from a known original, which makes it practical for training document restoration, OCR, and deskewing models. It also propagates masks, keypoints, and bounding boxes through spatial transforms, so it fits object-detection and segmentation workflows, not just image-to-image tasks. Teams that need realistic synthetic training data without collecting and labeling thousands of physical pages get the noisy copies and the clean originals in one package.