Pengenalan Suku Kata Aksara Bali Berbasis Aturan Pasang Aksara Dan Deep Learning Untuk Transliterasi Naskah Lontar Kuno

Sutramiani, Ni Putu (2022) Pengenalan Suku Kata Aksara Bali Berbasis Aturan Pasang Aksara Dan Deep Learning Untuk Transliterasi Naskah Lontar Kuno. Doctoral thesis, Institut Teknologi Sepuluh Nopember.

	Text 05111960010012-Disertation.pdf Restricted to Repository staff only Download (16MB) | Request a copy
	Text 05111960010012-Disertation.pdf - Accepted Version Restricted to Repository staff only until 1 April 2024. Download (16MB) | Request a copy

Abstract

Upaya pelestarian warisan budaya sangat penting dilakukan untuk menjaga warisan budaya dari kepunahan. Lontar merupakan dokumen kuno warisan budaya Bali yang berisikan sumber pengetahuan masyarakat Bali secara turun temurun. Sa-lah satu upaya konservasi naskah Lontar dilakukan dengan proses digitalisasi untuk menyelamatkan isi naskah yang terkandung dalam Lontar. Dibutuhkan pengetahuan mengenai aksara Bali khusus pada Lontar sehingga proses transliterasi naskah Lon-tar hanya dapat dilakukan oleh ahli sastra Bali. Hal tersebut melatarbelakangi penelitian ini untuk mengembangkan metode pengenalan suku kata aksara Bali pada Lontar kuno. Pengenalan suku kata Aksara Bali pada naskah Lontar memiliki tan-tangan yaitu (1) terbatasnya variasi yang mewakili penulisan aksara Bali pada Lon-tar, (2) aksara Bali pada Lontar memiliki variasi bentuk karakter dan noise pada citra Lontar karena proses pembuatan Lontar, (3) penulisan karakter aksara Bali memiliki keunikan dan berbeda dengan karakter latin yaitu adanya gantungan, gempelan, pengangge yang memiliki aturan tersendiri, (4) penulisan kalimat aksara Bali tidak memiliki spasi antar kata sehingga dibutuhkan aturan khusus pasang aksara Bali dalam mengenali suku kata aksara Bali.
Tujuan penelitian ini adalah membangun metode pengenalan suku kata aksara Bali yang efektif untuk transliterasi aksara Bali pada Lontar. Metode ini dibangun menggunakan aturan pasang aksara dan deep learning yang terdiri dari beberapa tahapan yaitu: (1) membangun dataset aksara Bali dengan menerapkan augmentasi data untuk menambah variasi dan peningkatan kualitas citra, (2) deteksi citra karakter aksara Bali yang padat dan variasi tinggi pada naskah Lontar menggunakan you only look once (YOLO) arsitektur versi YOLOv4, (3) pengenalan karakter aksara Bali pada naskah Lontar menggunakan arsitektur convolutional neu-ral network (CNN), dan (4) pengenalan suku kata aksara Bali berdasarkan aturan pasang aksara Bali.
Tahap pembangunan dataset menghasilkan dua dataset yaitu dataset Handwritten Balinese Characters of Lontar (HBCL) dan dataset untuk tugas de-teksi (HBCL_DETC). Dataset HBCL berisi kumpulan citra aksara Bali tulisan tan-gan pada Lontar yang terdiri dari 18 kelas aksara beserta hasil augmentasi menggunakan Multi Augmentation Technique – Adaptive Gaussian and Convolu-tional Autoencoder (MAT-AGCA) untuk pengembangan model pengenalan aksara Bali pada Lontar. Dataset HBCL_DETC berisi kumpulan citra naskah Lontar beser-ta hasil augmentasi yang dilengkapi dengan anotasi kelas dan posisi aksara untuk pengembangan model deteksi aksara Bali pada naskah Lontar. Pengenalan aksara Bali dibangun menggunakan pendekatan transfer learning pada beberapa model deep learning yaitu inception residual network V2 (InceptionResNetV2), dense con-volutional network (DenseNet169), residual network 152V2 (ResNet152V2), visual geometry group-16 (VGG19) dan MobileNetV2. Ujicoba menunjukkan bahwa model berbasis MobileNetV2 menghasilkan akurasi pengenalan terbaik, yaitu sebe-sar 96,29%. Pendeteksian aksara Bali pada naskah Lontar dibangun menggunakan YOLOv4 dan menghasilkan akurasi deteksi dengan mean average precision (mAP) sebesar 99,55%. Hasil deteksi aksara Bali pada naskah Lontar selanjutnya digunakan untuk mengenali suku kata aksara bali. pengenalan suku kata aksara Bali pada naskah Lontar dibangun berdasarkan posisi aksara dan penerapan aturan pasang aksara. Hasil eksperimen menunjukkan metode pengenalan suku kata aksara Bali dapat mengenali suku kata berdasarkan aturan pasang aksara dengan akurasi sebesar 71,18%.
=====================================================================================================
Efforts to preserve cultural heritage are significant to protect cultural heritage from extinction. Lontar is an ancient document of Balinese cultural heritage which contains a source of knowledge of the Balinese people from generation to generation. One of the efforts to conserve the Lontar manuscripts is carried out with a digitization process to save the contents of the manuscripts contained in the Lontar. It requires knowledge of the particular Balinese script in Lontar so that only Balinese literary experts could carry out the transliteration process of Lontar manuscripts. It became the background of this research to develop a method for recognizing Balinese script syllables in ancient Lontar. The recognition of Balinese script syllables in Lontar script had challenges, namely: (1) limited variations that represent the writing of Balinese script in Lontar, (2) Balinese script in Lontar has variations in character shape and noise in the Lontar image due to the process of making Lontar, (3) the writing of Balinese script characters is unique and different from Latin characters, namely the existence of gantungan, gempelan, pengangge which have their own rules, (4) the writing of Balinese script sentences does not have spaces between words, so it requires special rules for Balinese pasang aksara in recognizing Balinese script syllables.
The research aimed to develop an effective method of recognizing Balinese script syllables for transliterating Balinese script in Lontar. This method was built using the rules pasang aksara and deep learning, consisting of several stages. These were (1) building a Balinese script dataset by applying data augmentation to add variety and improve image quality, (2) detection of dense and highly variable Balinese character images in Lontar manuscripts using You Only Look Once (YOLO) version of the YOLOv4 architecture, (3) the recognition of Balinese characters in Lontar manuscripts using the Convolutional Neural Network (CNN) architecture, and (4) the recognition of Balinese script syllables based on the rules of pairs of Balinese characters.
The dataset development phase produced two datasets. These were the Hand-written Balinese Characters of Lontar (HBCL) dataset and the dataset for the detection task (HBCL_DETC). The HBCL dataset contained a collection of images of handwritten Balinese characters in Lontar consisting of 18 classes of characters and their augmentation results using the Multi Augmentation Technique – Adaptive Gaussian and Convolutional Autoencoder (MAT-AGCA) for the development of a model for recognizing Balinese characters in Lontar. The HBCL_DETC dataset contained a collection of Lontar manuscript images and augmentation results that were equipped with class annotations and characters positions to develop a Balinese character detection model in Lontar manuscripts. Balinese characters recognition was built using a transfer learning approach in several deep learning models, namely inception residual network V2 (Inception-ResNetV2), dense convolutional network (DenseNet169), residual network 152V2 (ResNet152V2), visual geometry group-16 (VGG19), and MobileNetV2. The experiment showed that the MobileNetV2-based model produced the best recognition accuracy of 96.29%. Balinese characters detection in Lontar manuscripts was built using YOLOv4 and produced a detection accuracy with a mean average precision (mAP) of 99.55%. Furthermore, the detection results of Balinese characters in Lontar manuscripts were used to identify Balinese script syllables. The recognition of Balinese script syllables in Lontar manuscripts was built based on the position of the characters and the application of the rules for pasang aksara. The experimental results showed that the Balinese script syllable recognition method could recognize syllables based on the rules of pasang aksara with an accuracy of 71.18%.

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	Balinese Characters, Deep Learning, Lontar Manuscripts, Noise, Pasang Aksara, Pengrupak, Cultural Heritage, Aksara Bali, Deep Learning, Naskah Lontar, Noise, Pasang Aksara, Pengrupak, Warisan Budaya.
Subjects:	Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55001-(S3) PhD Thesis (Comp Science)
Depositing User:	Ni Putu Sutramiani
Date Deposited:	01 Apr 2022 08:11
Last Modified:	01 Apr 2022 08:15
URI:	http://repository.its.ac.id/id/eprint/94879

Actions (login required)

View Item