Saturday 01 February 2025
The quest for a reliable and efficient Arabic handwritten text recognition system has been ongoing for years, with many researchers attempting to crack the code. A new study published in a recent journal article presents a promising solution that combines advanced machine learning techniques with traditional optical character recognition methods.
The team developed an Optical Character Recognition (OCR) system that uses a novel combination of differentiable binarization and adaptive scale fusion techniques for line segmentation, followed by a Convolutional Neural Network (CNN)-Bidirectional Long Short-Term Memory (BiLSTM)-Connectionist Temporal Classification (CTC) architecture for character recognition.
The researchers fine-tuned the Universal DBNet++ model on an Arabic dataset to improve its performance on handwritten text images. They used the Arabic Multi-Fonts Dataset (AMFDS), which contains 2 million word samples from 18 different fonts, with a maximum of 10 characters per word.
The study demonstrated that their OCR system achieved impressive results, with a Character Recognition Rate (CRR) of 99.20% and a Word Recognition Rate (WRR) of 93.75% on single-word samples containing 7-10 characters. The system also performed well on sentences, with a CRR of 83.76%.
One of the key challenges in Arabic handwritten text recognition is the cursive nature of the script, which can make it difficult to segment lines and recognize characters accurately. To address this issue, the researchers developed a novel line segmentation algorithm that uses differentiable binarization and adaptive scale fusion techniques.
The system first detects the lines in the image using the DBNet++ model, then applies differentiable binarization to extract features from each detected line. The adaptive scale fusion technique is used to combine these features and generate a probability map of the text lines.
The researchers also developed a novel OCR engine that uses a CNN-BiLSTM-CTC architecture to recognize characters in the segmented lines. The system first extracts features from each character using a convolutional neural network, then uses a bidirectional long short-term memory network to model the sequential relationships between characters.
Finally, the system applies connectionist temporal classification to align the input sequences with their corresponding labels and generate a decoded output.
The study demonstrates that the OCR system is highly accurate and efficient, even on challenging datasets. The researchers believe that their solution has the potential to revolutionize the field of Arabic handwritten text recognition, enabling more accurate and efficient digitization of documents and data extraction.
Cite this article: “Highly Accurate Arabic Handwritten Text Recognition System Developed Using Advanced Machine Learning Techniques”, The Science Archive, 2025.
Arabic, Handwritten, Text Recognition, Ocr, Machine Learning, Cnn, Bilstm, Ctc, Dbnet++, Line Segmentation, Character Recognition
