Metrics for Complete Evaluation of OCR Performance

Overview & Synopsis

In this paper, we study metrics for evaluating OCR performance both in terms of physical segmentation and in terms of textual content recognition. These metrics rely on the OCR output (hypothesis) and the reference (also called ground truth) input format. Two evaluation criteria are considered: the quality of segmentation and the character recognition rate. Three pairs of input formats are selected among two types of inputs: text only (text) and text with spatial information (xml). These pairs of inputs reference-to-hypothesis are: 1) text-to-text, 2) xml-to-xml and 3) text-to-xml. For the text-to-text pair, we selected the RETAS method to perform experiments and show its limits. Regarding text-to-xml, a new method based on unique word anchors is proposed to solve the problem of aligning texts with different information. We define the ZoneMapAltCnt metric for the xml-to-xml approach and show that it offers the most reliable and complete evaluation compared to the other two. Open source OCRs like Tesseract and OCRopus are selected to perform experiments. The datasets used are a collection of documents from the ISTEX 1 document database, from French newspaper "Le Nouvel Observateur" as well as invoices and administrative document gathered from different collaborations.