Document Denoising

Document denoising is the removal of visual noise from document images — the salt-and-pepper speckle of aging scans and faxes, sensor noise from low-light phone captures, scanner streaks and dust shadows, compression artifacts, show-through from a page's reverse side, and the physical marks of use: stains, smudges, hole punches, staple shadows. The goal is a cleaner separation between content and contamination, because noise costs recognition accuracy directly — speckle fragments characters or masquerades as punctuation, streaks cut through text lines, and artifacts inflate the false detections that downstream layout analysis must survive.

Techniques scale with noise complexity. Classical filters handle classical noise: median filtering for salt-and-pepper, morphological operations for isolated specks, Gaussian and bilateral smoothing (edge-preserving variants mattering for text, whose sharp edges are the signal). Frequency-domain methods remove periodic patterns like halftone screens from printed-photo regions. Learned denoising — encoder-decoder networks trained on noisy/clean document pairs, often synthetically corrupted — handles compound real-world degradation that no single filter models, and modern document-restoration models jointly denoise, deblur, and enhance in one pass. As with all document enhancement, the boundary discipline is conservatism: a denoiser aggressive enough to remove all speckle will eventually remove a decimal point, and a learned model can hallucinate clean-looking strokes — so critical-field regions warrant honest confidence rather than confidence inherited from a prettified image.

Pipeline placement is channel-aware: fax intake and archival scans earn heavy denoising; born-digital PDFs need none; phone captures need different treatment than flatbed scans. Mature systems profile noise per source, apply matched treatment, and validate the whole arrangement the only way that means anything — recognition accuracy on real documents, before and after, per channel.