Quality of OCR for Degraded Text Images

TL;DR

This paper investigates how the quality of OCR on degraded text images can be predicted using a simple noise model, aiding decisions on when human correction is necessary.

Contribution

It introduces a method to predict OCR failure rates on degraded images using controlled noise analysis, improving error estimation accuracy.

Findings

A simple noise model effectively predicts OCR errors on degraded images.

Controlled noise addition correlates with OCR failure rates.

Preliminary results show promising prediction accuracy.

Abstract

Commercial OCR packages work best with high-quality scanned images. They often produce poor results when the image is degraded, either because the original itself was poor quality, or because of excessive photocopying. The ability to predict the word failure rate of OCR from a statistical analysis of the image can help in making decisions in the trade-off between the success rate of OCR and the cost of human correction of errors. This paper describes an investigation of OCR of degraded text images using a standard OCR engine (Adobe Capture). The documents were selected from those in the archive at Los Alamos National Laboratory. By introducing noise in a controlled manner into perfect documents, we show how the quality of OCR can be predicted from the nature of the noise. The preliminary results show that a simple noise model can give good prediction of the number of OCR errors.