Comparing U-Net-ResNet34 and Modified Deep Semantic Binarization Network for Indonesian Ancient Document Segmentation

Zakharia, Rino (2026) Comparing U-Net-ResNet34 and Modified Deep Semantic Binarization Network for Indonesian Ancient Document Segmentation. Undergraduate thesis, UPN Veteran Jawa Timur.

	Text (cover) 21081010045-cover.pdf Download (795kB)
	Text (bab 1) 21081010045-bab1.pdf Download (59kB)
	Text (bab 2) 21081010045-bab2.pdf Restricted to Repository staff only until 22 July 2029. Download (329kB)
	Text (bab 3) 21081010045-bab3.pdf Restricted to Repository staff only until 22 July 2029. Download (1MB)
	Text (bab 4) 21081010045-bab4.pdf Restricted to Repository staff only until 22 July 2029. Download (5MB)
	Text (bab 5) 21081010045-bab5.pdf Download (7kB)
	Text (daftar pustaka) 21081010045-daftarpustaka.pdf Download (55kB)

Abstract

Indonesian ancient documents and manuscripts possess high historical value, but they often suffer from degradation that complicates the digitization process, particularly in the text and background binarization stage. This study aims to compare the performance of U-Net-ResNet34 and the Modified Deep Semantic Binarization Network (MDSBN) for Indonesian ancient document segmentation, while also using U-Net and ResNet34 as baseline models. The research stages include image preprocessing using grayscale conversion, median filtering, and CLAHE, image patch extraction, model training, performance evaluation, and computational requirement testing. The models were trained using BCEWithLogitsLoss and evaluated using Dice, IoU, precision, recall, RMSE, number of parameters, training time, inference time, and memory usage. The testing results based on micro average show that MDSBN slightly outperformed U-NetResNet34, achieving a Dice performance of 0.8545, IoU of 0.7459, precision value reaching 0.8600, recall performance 0.8490, and RMSE of 0.1529, while U-NetResNet34 achieved a Dice performance of 0.8535, IoU of 0.7445, precision value reaching 0.8581, recall performance 0.8490, and RMSE of 0.1535. Among the baseline models, U-Net achieved the highest Dice and IoU scores of 0.8559 and 0.7480, respectively, whereas ResNet34 achieved the lowest performance with a Dice performance of 0.3156 and an IoU value 0.1874. In terms of computational requirements, MDSBN has fewer parameters, totaling 15,327,681 compared to 24,349,441 for U-Net-ResNet34, but U-Net-ResNet34 has a faster inference time of 2.38 ms per image compared to 9.08 ms per image for MDSBN. Based on these results, MDSBN can serve as a competitive alternative with a smaller number of parameters, while U-Net-ResNet34 is superior in terms of inference efficiency

Item Type:

Thesis (Undergraduate)

Contributors:

Contribution	Contributors	NIDN/NIDK	Email
Thesis advisor	Nugroho, Budi	NIDN0707098003	budinugroho.if@upnjatim.ac.id
Thesis advisor	Mandyartha, Eka Prakarsa	NIDN0725058805	eka_prakarsa.fik@upnjatim.ac.id

Subjects:

Q Science > QA Mathematics > QA76.87 Neural computers

Divisions:

Faculty of Computer Science > Departemen of Informatics

Depositing User:

Rino Zakharia

Date Deposited:

22 Jul 2026 08:27

Last Modified:

23 Jul 2026 01:32

URI:

https://repository.upnjatim.ac.id/id/eprint/57802