Empirical Evaluation of Well-known Farsi OCR Engines on the IDPL-PFOD Dataset

Optical character recognition (OCR), also referred to as text recognition, extracts text from scanned documents and camera images. OCR has numerous applications in reading forms and cheques, converting archived documents to digital files, and reading books and papers. An accurate OCR system speeds up these processes by eliminating time-consuming manual tasks. Despite advancements, OCR remains challenging for languages like Farsi due to their complex structures and lack of diverse datasets for evaluation. IDPL-PFOD, a new synthetic Farsi dataset, addresses this gap by including various backgrounds, fonts, distortions, and blurs. This paper evaluates Tesseract and EasyOCR on IDPL-PFOD to highlight their limitations. Tesseract and EasyOCR achieve ۸۴.۴۱% and ۷۳.۲۸% overall accuracy, respectively. Tesseract outperforms EasyOCR in most cases but drops to ۶۷.۱۷% on textured backgrounds, while EasyOCR achieves ۷۵.۸%. Tesseract reaches its peak accuracy (۹۲.۴۵%) on plain backgrounds and shows robustness to Gaussian blur and slope distortion, with minimal accuracy drops. In contrast, EasyOCR performs poorly on salt-and-pepper noise (۴.۷۹%), and moderately on sinewave distortion and Gaussian blur (۶۱.۴۵% and ۶۳.۷۳%). These findings emphasize the need for training OCR engines on more diverse datasets to improve their performance under real-world conditions.

کلیدواژه ها:

IDPL-PFOD dataset ، optical character recognition (OCR) ، EasyOCR ، Tesseract ، OCR engines ، Farsi OCR