StructFusionNet: Structural Feature Fusion in CNNs for Handwritten

Babylonian numerals constitute a fundamental component of the Mesopotamian sexagesimal numbering system and represent one of the earliest structured numeral representations in human history. Despite their historical importance, the automated recognition of Babylonian numeral symbols remains relatively underexplored in computational research. This study proposes StructFusionNet, a convolutional neural network (CNN) framework that integrates learned visual representations with handcrafted structural features for the classification of Babylonian numeral forms. To evaluate the proposed approach under realistic handwriting variability, a dataset consisting of 14,000 segmented images representing 14 Babylonian numeral classes (1–9, 10, 20, 30, 40, and 50) was constructed from handwritten samples collected from 100 participants using structured forms. The proposed framework combines convolutional feature extraction with structural descriptors derived from edge detection, line detection, and corner analysis, enabling improved discrimination between visually similar numeral symbols. Experimental results demonstrate that the proposed StructFusionNet model achieves an overall classification accuracy of 97.39%, with precision, recall, and F1‐score values of 97.39%, 97.44%, and 97.40%, respectively. An ablation study further confirms that integrating structural geometric descriptors significantly improves classification performance compared with baseline CNN models and edge‐only feature variants. It is important to note that the dataset consists of modern handwritten reproductions of Babylonian numeral forms rather than photographs of archaeological clay‐tablet inscriptions. Therefore, the proposed system should be interpreted as a controlled proof‐of‐concept for handwritten Babylonian numeral classification, providing a reproducible benchmark that may support future research involving scholarly handcopies and real cuneiform tablet imagery.