Quranic diacritic and character segmentation and recognition using flood fill and k-nearest neighbors algorithm

The detection, recognition and conversion of the characters in an image into a text are called optical character recognition (OCR). A distinctive type of OCR is used to process Arabic characters, namely, Arabic Optical Character Recognition (AOCR). OCR is increasingly used in many applications, where this process is preferred to automatically perform a process without human intervention. The Quranic handwriting text contains two elements, namely, diacritics and characters. However, the current Arabic handwritten OCR system produces low levels of accuracy and no research focused on Quran image recognition. The current AOCR inaccurately recognizes diacritic and characters, and the research and efforts in the area of AOCR are insufficient. Many studies have been carried out so far, but for Quran handwriting has not been researched as thoroughly as Arabic, Latin or Chinese handwritten systems. The current research is focused on solving the mentioned problems through improving the accuracy of recognition rate of AOCR by proposing a new segmentation, feature extraction methods and finding a suitable classification. In this thesis, a new techniques, methods and algorithms are proposed to check the similarities and originalities of the Quranic handwriting content. The diacritic detections are performed using a region-based algorithm with 89% accuracy and 95% improved by using flood fill segmentations method. 2DMED feature extraction accuracy was 90% for diacritics and 96% improved by applied CNN. Character recognition is performed based on the projection method with 86% accuracy, and 92% improved by using flood fill. 2DMED in characters was 88% and 91 % after improved by applied CNN. For classification, KNN used before and after enhancement technique based on essential vector with our dataset, the diacritic accuracy was 96.4286% after enhancement, which is better than the 87.5020% in detecting. For characters was at 92.3077% improvement, which is better that normal KNN algorithm which exhibited an 86.1429% accuracy in detecting.

Text FSKTM 2019 59 IR.pdf Download (1MB)

Actions (login required)

View Item