Novel Deep Convolutional Neural Network-Based Contextual Recognition of Arabic Handwritten Scripts.

Area of Science:

• Pattern Recognition
• Image Processing
• Deep Learning

Background:

• Offline Arabic Handwriting Recognition (OAHR) is crucial for document processing but faces challenges due to script variability and lack of data.
• Existing deep learning models struggle with the complexities of Arabic script and human handwriting variations.

Purpose of the Study:

• To develop a novel context-aware deep neural network model for accurate OAHR.
• To address challenges in recognizing isolated digits, characters, and words in handwritten Arabic text.

Main Methods:

• A supervised Deep Convolutional Neural Network (DCNN) architecture with stacked convolutional layers was proposed.
• Batch normalization and dropout regularization were employed to prevent overfitting and enhance generalization.
• Transfer learning was utilized for feature extraction, comparing performance against pre-trained models like VGGNet-19 and MobileNet.

Main Results:

• The proposed DCNN model demonstrated excellent classification accuracy on six benchmark Arabic databases.
• The model outperformed conventional OAHR approaches and state-of-the-art pre-trained models in transfer learning experiments.
• Comparative analysis on the MNIST English digits database confirmed the DCNN model's superior generalization capabilities.

Conclusions:

• The novel context-aware DCNN model effectively addresses the challenges of OAHR.
• The proposed architecture offers superior performance and generalization compared to existing methods.
• This work advances the field of handwritten Arabic text recognition, with potential applications in automation and document processing.