Issues in automatic OCR error classification
Jeffrey Esakov, Daniel Lopresti, Jonathan S. Sandberg, Jie Zhou · OSTI OAI (U.S. Department of Energy Office of Scientific and Technical Information) · 1994
In this paper we present the surprising result that OCR errors are not always uniformly distributed across a page. Under certain circumstances, 30% or more of the errors incurred can be attributed to a single, avoidable phenomenon in the scanning process. This observation has important ramifications for work that explicitly or implicitly assumes a uniform error distribution. In addition, our experiments show that not just the quantity but also the nature of the errors is affected. This could have an impact on strategies used for post-process error correction. Results such as these can be obtained only by analyzing large quantities of data in a controlled way. To this end, we also describe our algorithm for classifying OCR errors. This is based on a well-known dynamic programming approach for determining string edit distance which we have extended to handle the types of character segmentation errors inherent to OCR.