Handcrafted and Deep Feature Analysis for Computer-Aided Dysgraphia Diagnosis from Handwriting Images

Dysgraphia is a neurodevelopmental disorder that affects handwriting and written expression. While most automated detection systems rely on online handwriting captured by digitizing tablets, these approaches face scalability and hardware constraints. Offline analysis from scanned handwriting images provides a more practical and accessible alternative. This study systematically compares classical handcrafted feature, HOG, LBP, Gabor statistics, Hu moments, density, and entropy, with modern CNNs including ResNet18, EfficientNet-B0, and a lightweight SmallCNN, all evaluated under identical stratified cross-validation protocols. Results show that handcrafted feature pipelines achieved up to 96% best-fold accuracy (mean 88–89%) using SVM and Logistic Regression, outperforming deep CNNs. Compared to prior studies on the same dataset, the proposed offline framework achieves comparable or superior performance with strong statistical reliability. These findings demonstrate that handcrafted descriptors remain competitive for interpretable and resource-efficient dysgraphia screening, supporting practical diagnostic tools in low-resource and educational contexts.