<![CDATA[The digitization of ancient documents is an important step in preserving historical and cultural information. However, the resulting images often suffer from degradation, such as stains, uneven background textures, faded ink, and low contrast, making text-background separation difficult. This study compares two deep learning architectures, namely U-Net ResNet34 and the Modified Deep Semantic Binarization Network (MDSBN), for the segmentation of Indonesian ancient documents. The dataset consists of Balinese palm-leaf manuscripts, Sundanese manuscripts, and additional ancient document images obtained from Wikimedia Commons. The experiments were conducted through a learning rate search and batch size sensitivity analysis, and the models were evaluated using the Dice Coefficient, Intersection over Union (IoU), Precision, Recall, and Root Mean Squared Error (RMSE). This study contributes through a controlled evaluation of both architectures using a consistent dataset, preprocessing pipeline, loss function, evaluation metrics, and computational environment, enabling performance differences to be analyzed more objectively. The results show that U-Net ResNet34 achieved its best performance using a learning rate of 5e-5 and a batch size of 16, with a test Dice score of 0.79338 and a test IoU score of 0.65752. It outperformed MDSBN, which achieved its best performance using a learning rate of 1e-6 and a batch size of 32, with a test Dice score of 0.75338 and a test IoU score of 0.60433. The functional advantage of U-Net ResNet34 is associated with the ability of its residual encoder to extract hierarchical features from complex textures and degradation patterns, while its skip connections help preserve the spatial details of thin text strokes. These characteristics make U-Net ResNet34 more adaptive to variations in degradation within the Indonesian ancient document dataset than the more compact MDSBN architecture.]]>
Copyrights © 2026
