Open Vocabulary Recognition of Offline Arabic Handwriting Text Based on Deep Learning

Using Faster R-CNN to Detect and Recognize Arabic Handwritten Words

“…A few studies are available to identify unconstrained Arabic text with an extensive vocabulary. Recognizing Arabic text is challenging because of the various features of this language [3,4]. These challenges include the complexity that arises because of the cursive nature of the script, the connectivity between characters, the diverse styles of writers, extensive vocabulary, the presence of ligatures, overlaps, and irregular spacing [1,3] The handwriting recognition system's principal goal is to convert handwritten text documents from digital image format into encoded character format documents so that programs for word processing can read and change them.…”

“…International Journal of Intelligent Engineering and Systems, Vol.17, No. 3 Recently, researchers have endeavoured to develop methods for recognizing Arabic handwritten words using the IFN/ENIT. In 2021, Maalej and Kherallah [12] suggested an offline Arabic handwriting recognizer that used a Multi-Dimensional Long Short-Term Memory Network (MDLSTM) and Rectified Linear Units (ReLUs) to fix issues with vanishing gradients and dropout to avoid overfitting.…”

“…They achieved an accuracy rate of 80.02% for character recognition, a character error rate of 4.22%, and a label error rate of 19.98% in the KHATT dataset. In addition, in 2020, Zouhaira et al [3] Suggested open-vocabulary offline recognition method for handwritten Arabic text was developed, focusing on enhancing image quality through preprocessing steps. The approach used a deep Convolutional Recurrent Neural Network (CRNN) model with a VGGNet architecture and a BLSTM layer with CTC beam search decoder.…”

“…The authors in [12] incorporated the Rectified Linear Unit (ReLU) to address this challenge. To reduce overfitting-related issues, [3,12,13] used regularization techniques like dropout. During the feature extraction step, [1,3] employed a sliding window approach to feeding data into the CNN, which involves scanning the original image horizontally using a sliding window that moves from right to left.…”

“…To reduce overfitting-related issues, [3,12,13] used regularization techniques like dropout. During the feature extraction step, [1,3] employed a sliding window approach to feeding data into the CNN, which involves scanning the original image horizontally using a sliding window that moves from right to left. The dimensions of the window matched the width of the input text-line images.…”