Rekonstruksi Aksara Jawa Pada Naskah Kuno Menggunakan Deep Learning

Damayanti, Fitri (2025) Rekonstruksi Aksara Jawa Pada Naskah Kuno Menggunakan Deep Learning. Doctoral thesis, Institut Teknologi Sepuluh Nopember.

Text 07111960010017-Dissertation.pdf - Accepted Version Restricted to Repository staff only until 1 April 2027. Download (28MB) | Request a copy

Abstract

Pelestarian aksara Jawa pada naskah kuno sangat penting untuk menjaga warisan budaya yang kaya nilai historis. Namun, kerusakan fisik seperti tinta memudar, robekan, bercak noda, dan tinta tembus dari halaman sebelumnya menyebabkan hilangnya informasi berharga dalam naskah. Penelitian ini bertujuan untuk merekonstruksi aksara Jawa menggunakan pendekatan Deep Learning dengan fokus pada tiga tahap utama: segmentasi foreground-background, deteksi aksara, dan rekonstruksi aksara yang rusak.
Pada tahap segmentasi foreground-background, model ResUNet yang dimodifikasi dan dikombinasikan dengan Convolutional Long Short-Term Memory (ConvLSTM) digunakan untuk memisahkan area aksara (foreground ) dari latar belakang yang terdegradasi (tinta tembus dan terdapat bercak) serta berlubang. ResUNet dipilih karena kemampuannya dalam menangkap fitur multi-skala melalui blok residual, sementara ConvLSTM digunakan untuk mempelajari hubungan spasial dan temporal yang kompleks pada area naskah yang mengalami degradasi. Hasil evaluasi menunjukkan nilai loss sebesar 0,0559, F-Measure sebesar 92,89%, PSNR sebesar 18,52 dB, dan IoU sebesar 0,85.
Tahap deteksi aksara menggunakan metode Connected Component Labeling (CCL) untuk memisahkan aksara satu per satu. Metode ini menghasilkan akurasi rata-rata sebesar 80,9% dalam mendeteksi dan memisahkan aksara Jawa tulisan tangan dan cetak dengan variasi bentuk yang kompleks. Pada tahap rekonstruksi aksara, teknik character inpainting berbasis Deep Learning diterapkan menggunakan arsitektur U-Net, ResUNet, Partial Convolutional Neural Network (PCNN), dan Convolutional Autoencoder (CAE). Hasil evaluasi menunjukkan bahwa model ResUNet memberikan performa terbaik dengan nilai
PSNR sebesar 18,95 dB dan SSIM sebesar 0,9319 dibandingkan dengan metode lainnya.
Penelitian ini berhasil mengatasi tantangan dalam pemulihan aksara Jawa yang terdegradasi, sehingga diharapkan dapat mendukung upaya revitalisasi warisan budaya dan menjadi referensi dalam pengembangan metode pemrosesan citra naskah kuno.
====================================================================================================================================
The preservation of the Javanese script in ancient manuscripts is very important to preserve a cultural heritage rich in historical value. However, physical damages such as ink fading, tears, stains, and ink bleeding from previous pages lead to the loss of valuable information in the manuscripts. This research aims to reconstruct Javanese script using a deep learning approach by focusing on three main stages: foreground-background segmentation, segmentation of individual characters (each script), and reconstruction of damaged scripts. In the foreground-background segmentation stage, a modified ResUNet model combined with Convolutional Long Short-Term Memory (ConvLSTM) is used to separate the script area (foreground) from the degraded background (translucent ink and stains) and holes. ResUNet was chosen for its ability to capture multi-scale features from residual blocks, while ConvLSTM was used to learn complex spatial and temporal relationships in degraded script areas. The evaluation results show a loss value of 0.0559, an F-measure of 92.89%, a PSNR of 18.52 dB, and an IoU of 0.85. The single character segmentation stage uses the Connected Component Labeling (CCL) method to separate the characters one by one. This method achieves an average accuracy of 80.9% in detecting and separating handwritten and printed Javanese characters with complex shape variations. In the script reconstruction stage, deep learning-based character inpainting techniques were applied using U-Net, ResUNet, Partial Convolutional Neural Network (PCNN), and Convolutional Autoencoder (CAE) architectures. The evaluation results show that the ResUNet model provides the best performance with a PSNR value of 18.95 dB and SSIM of 0.9319 compared to other methods. This research successfully overcomes the challenges in the recovery of degraded Javanese script, so it is expected to support cultural heritage revitalization efforts and become a reference in the development of ancient manuscript image processing.

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	Javanese Script, Deep Learning, Ancient Manuscript, Foreground- Background Segmentation, Character Segmentation, Reconstruction,Aksara Jawa, Deep Learning, Naskah Kuno, Segmentasi Foreground-Background, Segmentasi Karakter, Rekonstruksi
Subjects:	R Medicine > R Medicine (General) > R858 Deep Learning T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7882.P3 Pattern recognition systems
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Electrical Engineering > 20001-(S3) PhD Thesis
Depositing User:	Fitri Damayanti
Date Deposited:	03 Feb 2025 01:59
Last Modified:	03 Feb 2025 01:59
URI:	http://repository.its.ac.id/id/eprint/117791

Actions (login required)

View Item