Handwritten Javanese Script Text Recognition Using CNN BiLSTM and Connectionist Temporal Classification

Prastya, Ade Fathoni (2026) Handwritten Javanese Script Text Recognition Using CNN BiLSTM and Connectionist Temporal Classification. Undergraduate thesis, UPN Veteran Jawa Timur.

	Text (Cover) 22081010204.-cover.pdf Download (2MB)
	Text (Bab 1) 22081010204.-bab1.pdf Download (193kB)
	Text (Bab 2) 22081010204.-bab2.pdf Restricted to Repository staff only until 10 June 2028. Download (984kB)
	Text (Bab 3) 22081010204.-bab3.pdf Restricted to Repository staff only until 10 June 2028. Download (1MB)
	Text (Bab 4) 22081010204.-bab4.pdf Restricted to Repository staff only until 10 June 2028. Download (1MB)
	Text (Bab 5) 22081010204.-bab5.pdf Download (142kB)
	Text (Daftar pustaka) 22081010204.-daftarpustaka.pdf Download (187kB)
	Text (Lampiran) 22081010204.-lampiran.pdf Restricted to Repository staff only Download (338kB)

Abstract

The Nglegena Javanese script is a traditional writing system with high historical and cultural value, but faces the threat of extinction due to the low literacy skills of the community. Based on previous research, 81.4% of students are unable to read and 79% are unable to write Javanese script. On the other hand, more than 19,000 Javanese script manuscripts stored in various collections have not been adequately inventoried, so an automatic text recognition system is needed that can support digitization and preservation efforts. This study aims to develop an end-to-end handwritten text recognition model for the Nglegena Javanese script based on Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory (BiLSTM), and Connectionist Temporal Classification (CTC) architecture without requiring explicit character segmentation. The dataset used consists of 1,200 synthetic images resulting from rendering Javanese script fonts, with 1,000 images as training data and 200 images as validation data. In addition, 300 real handwritten images from six participants were used, divided into 100 images as validation data and 200 images as test data. This study systematically explored architectural variations by varying the number of CNN layers from 3 to 7 layers and BiLSTM layers from 1 to 3 layers, using Character Error Rate (CER) and Exact Match (EM) as evaluation metrics. The experimental results showed that the optimal configuration was achieved by the 5-CNN and 2-BiLSTM architectures with a CER value of 0.068 and an EM accuracy of 0.795. Fine architectures (3-4 CNN layers) indicated underfitting due to limited feature capacity, while deeper architectures (6-7 CNN layers and 3 BiLSTM layers) showed performance degradation and instability due to overfitting. The effectiveness of the CNN-BiLSTM-CTC architecture synergy was proven through the model's high generalization ability on varied real handwritten test data. The entire system is then implemented in the form of an API and an interactive website interface to support accessibility of Nglegena Javanese script recognition by other systems and the wider community.

Item Type:

Thesis (Undergraduate)

Contributors:

Contribution	Contributors	NIDN/NIDK	Email
Thesis advisor	Anggraeny, Fetty Tri	NIDN0711028201	fettyanggraeny.if@upnjatim.ac.id
Thesis advisor	Puspaningrum, Eva Yulia	NIDN0005078908	evapuspaningrum.if@upnjatim.ac.id

Subjects:

Q Science > Q Science (General)
Q Science > QA Mathematics > QA76.6 Computer Programming
Q Science > QA Mathematics > QA76.87 Neural computers

Divisions:

Faculty of Computer Science > Departemen of Informatics

Depositing User:

Ade Fathoni Prastya

Date Deposited:

11 Jun 2026 07:02

Last Modified:

11 Jun 2026 07:02

URI:

https://repository.upnjatim.ac.id/id/eprint/53889