Assessing the Relationship Between Binarization and OCR in the Context of Deep Learning-Based ID Document Analysis

• Published in: Progress in Artificial Intelligence and Pattern Recognition pp. 134 - 144
• Main Authors: Sánchez-Rivero, Rubén, Bezmaternykh, Pavel, Morales-González, Annette, Silva-Mata, Francisco José, Bulatov, Konstantin
• Format: Book Chapter
• Language: English
• Published: Cham Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Deep learning; Document image binarization; Identity document recognition; Optical character recognition
• ISBN: 3030896900; 9783030896904
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-030-89691-1_14

Text recognition has been one of the areas greatly benefited from deep learning (DL) development, as well as the preprocessing methods contained in its workflow. Within the document analysis field, identity (ID) documents play a crucial role and should be studied in depth regarding this workflow. For this reason, we propose to analyze the relationship between DL-based binarization and recognition methods, specifically for this type of documents. We perform a review of four binarization and seven optical character recognition (OCR) algorithms, and present two sets of experiments assessing the influence of text size for binarization, and the impact of its output in the final text recognition. We show that DL-based binarization solutions are very sensitive to logical text size and they are still not effective in this domain, thus requiring major improvements. Among the evaluated text recognizers, the best performance over the binarization results was shown by the semantic reasoning network (SRN) method.