GreekNet: Handwritten Greek Alphabet Recognition Using Explainable Parallel CNN with Attention Mechanisms
Anwar Hossain Efat, S. M. Mahedy Hasan, Minhaz F. Zibran
Abstract
The Greek alphabet is one of the most ancient and influential writing systems, playing a crucial role in disciplines such as mathematics, physics, and engineering. Despite its significance, accurately recognizing handwritten Greek characters remains a challenge due to variations in individual handwriting and the absence of reliable techniques capable of handling these challenges. To overcome these challenges, we have first developed a dataset comprising 6,015 samples of lowercase Greek characters, collected from around 300 individuals to ensure a comprehensive representation of handwriting styles. Then we propose a novel machine learning (ML) architecture, GreekNet, that uses this dataset and combines a customized DenseNet121 with three parallel Custom Convolutional Neural Networks (CNN), each integrating different attention mechanisms: Channel Attention (CA), Squeeze-Excitation Attention (SEA), and Soft Attention (SA). These attention modules operate in parallel, with their outputs concatenated to generate more complex and distinguishable feature maps. Our proposed architecture achieved 99.35% accuracy. Using a Grad-CAM visualization approach we achieve explainability of our ML approach.