Multiclass dysgraphia severity classification in Malay handwriting using enhanced image processing techniques and feature-fusion based machine learning

Dysgraphia is a learning disability (LD) that causes difficulties in writing and commonly occurs among school-age children. One of its key symptoms is poor, inconsistent handwriting, typically identified through expert observation with prescribed scoring systems. However, such conventional assessments are time-consuming, prone to interruption, and may produce subjective or biased results. With advances in technology, automated dysgraphia detection methods have emerged, leading to various handwriting-based approaches. However, most existing studies emphasize dynamic handwriting features, such as tilt, angle, pressure level, and pen position, and are limited to binary classification (dysgraphia vs. non-dysgraphia). Consequently, multiclass severity classification using offline handwritten images remains underexplore. To address this gap, this study proposes a novel multiclass classification approach for assessing dysgraphia severity using offline Malay handwriting samples collected from children aged six to twelve years. A new dataset, the Malay Language Dysgraphia Handwriting Dataset (MLHSD) was developed to capture diverse handwriting characteristics. A robust preprocessing scheme based on a Red-Green-Blue-Black (RGBB) ruling-line removal algorithm was introduced to enhance image quality while preserving essential handwriting features. Subsequently, the Dual-Input Feature Fusion Model (H2FCD) was developed, integrating feature extraction from two complementary inputs: resized handwritten images with removed ruling lines (HIRRL) and histogram-based representations (Histo-HIRRL). These fused feature vectors were then classified across multiple machine-learning classifiers within a multiclass classification. Experimental results showed that the Support Vector Machine (SVM) achieved the best performance, with an accuracy of 99.26%, precision of 99.49%, recall of 99.52%, and F1-score of 99.51%. The robustness of the proposed model was further validated using unseen test data, achieving an accuracy of 88.89% and Matthews Correlation Coefficient (MCC) of 0.839, indicating strong agreement between predicted and actual classes. Additionally, expert validation by eight handwriting specialists yielded a moderately strong correlation coefficient of 0.640 between the model outputs and expert evaluations. These findings demonstrate that the proposed H2FCD model, combined with RGBB-based preprocessing and dual-input feature fusion, significantly improves multiclass classification of dysgraphia severity using offline Malay handwritten images, contributing to more reliable and objective screening tools for early intervention.