A Printed Arabic Optical Character Recognition System using Deep Learning

[摘要] RecognizingArabic script is challenging for many reasons: The Arabic language is cursiveand morphologically rich. There is a high similarity between Arabic letters.Moreover, Arabic has many diacritics and dots, and they change the letter'sphonetic transcription. This study proposes a Printed Arabic Optical CharacterRecognition approach (PAOCR) based on the state-of-the-art You Only Look Once(YOLO) object detector. Four techniques were proposed and implemented to designan end-to-end Arabic OCR system. First, the YOLO4 object detector is customizedand trained on deep Convolutional Neural Networks (CNNs) to recognize Arabiccharacters. Second, the overlapped bounding boxes are processed to keep themost accurate box for each character. Third, the Hunspell library checks theword spelling and corrects the wrong ones. Fourth, edit distance is used tocompare OCR misspelled words with Hunspell's suggestions and choose the closestcorrect word. The experimental results in the Arabic Printed Text Image (APTI)dataset showed that the Word Recognition Rate (WRR) of customized YOLO4 ArabicOCR is 66.7%, whereas applying the three proposed techniques with YOLO4achieved 82.4%. A comparison with two existing OCR systems Tesseract and ABBYYshowed that the proposed approach achieved a 95.7% Character Recognition Rate(CRR), whereas Tesseract and ABBYY achieved 92.3 and 84.8%, respectively.

[学科分类] 计算机科学（综合）

[关键词] Optical Character Recognition;Arabic OCR;Convolutional Neural Network;YOLO4