Abstract
The Ge'ez script is one of the ancient writing systems. It is underrepresented in modern OCR research due to its complex numeral shapes and the absence of standardized datasets. This paper presents GeezNet, a deep learning framework optimized for Ge'ez numeral recognition. We also introduce a curated, open-access dataset containing 10,000 labeled images across 20 numeral classes. Several GeezNet variants-baseline, depthwise separable, residual, and Vision Transformer (ViT) hybrid-are compared with transfer learning models such as ResNet-50 and Inception-V3. The optimized 15-layer GeezNet achieves 97.8% accuracy on an independent test set. It reduces computational cost compared to larger generic architectures, enabling deployment on mobile and archival digitization systems. Misclassification analysis shows difficulties with visually similar numerals. These insights inform future work on robustness to noise and handwriting variability. This study provides a benchmark model, a comprehensive architecture comparison, and a public dataset to advance OCR for underrepresented scripts in low-resource settings.