Image and text correction using language models

We report on experiments with the use of learned classifiers for improving OCR accuracy and generating word-level correction candidates. The method involves the simultaneous application of several image- and text-correction models, followed by a performance evaluation that enables the selection of the best image-processing model for each document and the most likely corrections for each word. It relies on a training set comprising document images and their transcriptions, plus a domain corpus used to build the language model. It is applicable to any language with simple segmentation rules and performs well on morphologically-rich languages. Experiments with an Arabic newspaper corpus show a 50% reduction in word error rate, with per-document image enhancement a major contributor.