OCRnoise

OCRnoise is a term used to describe the degradation and distortions in document images or recognition processes that reduce the accuracy of optical character recognition (OCR) systems. It encompasses image-level noise, capture artifacts, and processing artifacts that hinder reliable text extraction.

Common sources include scanner and camera noise, motion blur, uneven illumination, paper texture, ink bleed, stains, compression artifacts, and binarization errors. Noise can appear as pixel-level random variations, blur, speckle, or stray marks that disrupt character shape and layout.

OCRnoise increases character, word, and line recognition errors, leading to higher character error rate (CER) and word error rate (WER). It also degrades segmentation, line detection, and layout analysis, and lowers OCR confidence scores, often triggering post-processing corrections.

Metrics used to assess impact include CER and WER. Image quality metrics such as PSNR or SSIM quantify noise levels but do not directly reflect OCR performance. Researchers evaluate robustness by comparing OCR accuracy on datasets with varying noise levels or simulated distortions.

Mitigation includes preprocessing and denoising (median, Wiener, non-local means), adaptive thresholding, contrast enhancement, deblurring, and deskewing. Training OCR models on noisy data, data augmentation, and applying language models or post-correction improve resilience to OCRnoise.

Understanding OCRnoise aids digitization projects, archival preservation, and tool benchmarking, helping engineers select preprocessing pipelines and OCR systems appropriate to image quality and capture conditions.