Handwritten text segmentation using deep learning method

Abstract

The rapid development of artificial intelligence and deep learning technologies has increased the risk of digital task fabrication in academic environments, encouraging educators to reintroduce handwritten assignments as an authentic evaluation method. In handwritten document analysis systems, background segmentation is a critical preprocessing step that separates text from complex document backgrounds. This study proposes the use of the U-Net deep learning architecture for background segmentation of handwritten document images. Two datasets were employed: the public cBAD dataset and a custom dataset consisting of Indonesian handwritten student assignments. Both datasets were processed using an identical pipeline and evaluated using 5-fold cross-validation. Model performance was measured using the Dice Similarity Coefficient and Intersection over Union (IoU). Experimental results show that the proposed U-Net model achieved an average Dice coefficient (F1-Score) of 0.74 on the cBAD dataset and 0.83 on the student assignment dataset. These results indicate that the model performs consistently and demonstrates stable generalization across cross-validation folds. Therefore, the proposed approach is suitable as an initial segmentation stage in handwritten document recognition systems.