Pengembangan Model BLIP untuk Pembangkitan Keterangan Otomatis pada Citra Seni Ukiran Bali

Prasetyo Raharja, I Putu Bagus Gede (2026) Pengembangan Model BLIP untuk Pembangkitan Keterangan Otomatis pada Citra Seni Ukiran Bali. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025231010_Master Thesis.pdf Restricted to Repository staff only Download (20MB) | Request a copy

Abstract

Ukiran Bali merupakan salah satu warisan budaya Indonesia yang menonjol dengan ragam hias khas dan nilai-nilai simbolis yang terkandung dalam setiap motifnya. Ukiran ini menghiasi berbagai bangunan tradisional seperti pura dan rumah adat, mencerminkan kekayaan budaya dan filosofi hidup masyarakat Bali. Namun, pelestarian ukiran Bali menghadapi tantangan signifikan di tengah arus globalisasi dan modernisasi yang mengancam keaslian bentuk dan teknik ukirannya. Upaya digitalisasi budaya Bali khususnya ukiran Bali terkendala oleh keterbatasan data yang tersedia dan kompleksitas keterkaitan antar berbagai motif yang membentuk struktur kombinasi unik. Penelitian ini bertujuan untuk mengatasi hambatan tersebut melalui pengembangan metode pembangkitan keterangan citra otomatis berbasis model Bootstrapping Language-Image Pre-training (BLIP). BLIP merupakan arsitektur vision-language yang telah dilatih pada jutaan pasangan citra-teks, memungkinkan pemahaman mendalam terhadap konten visual dan kemampuan menghasilkan keterangan tekstual yang koheren. Namun, model pra-latih seperti BLIP memerlukan adaptasi khusus agar dapat mengenali karakteristik unik domain budaya lokal. Penelitian ini mengembangkan Multi-Label Cultural-Aware Adapter BLIP (ML-CAA-BLIP), sebuah pendekatan adaptasi efisien dengan menyisipkan modul adaptor ringan ke dalam arsitektur BLIP. Metode ini mengintegrasikan mekanisme culture aware bias dengan representasi visual berdasarkan konteks kelas budaya. Pengujian dilakukan menggunakan dataset BaliCarving yang terdiri dari 2.181 citra ukiran Bali dengan 7 kelas budaya. Hasil evaluasi menunjukkan ML-CAA-BLIP mencapai skor BLEU-4 sebesar 0,2718 dan ROUGE-L sebesar 0,5835, meningkat masing-masing 52,4% dan 15,2% dibandingkan model BLIP dasar.
===========================================================================================================================
Balinese carving is one of Indonesia's prominent cultural heritages, distinguished by its unique decorative patterns and symbolic values embedded in each motif. These carvings adorn various traditional buildings such as temples and traditional houses, reflecting the cultural richness and life philosophy of Balinese society. However, the preservation of Balinese carving faces significant challenges amid globalization and modernization, which threaten the authenticity of its forms and techniques. Efforts to digitize Balinese culture, particularly Balinese carvings, are constrained by limited available data and the complexity of interrelationships among various motifs that form unique combinatorial structures. This research aims to address these obstacles through the development of an automatic image captioning method based on the Bootstrapping Language-Image Pre-training (BLIP) model. BLIP is a vision-language architecture pre-trained on millions of image-text pairs, enabling deep understanding of visual content and the ability to generate coherent textual descriptions. However, pre-trained models like BLIP require specialized adaptation to recognize the unique characteristics of local cultural domains. This research develops Multi-Label Cultural-Aware Adapter BLIP (ML-CAA-BLIP), a parameter-efficient adaptation approach that inserts lightweight adapter modules into the BLIP architecture. This method integrates a Cultural-Aware Bias mechanism with visual representations based on cultural class context. Testing was conducted using the BaliCarving dataset consisting of 2,181 images of Balinese carvings across 7 cultural classes. Evaluation results demonstrate that ML-CAA-BLIP achieves a BLEU-4 score of 0.2718 and ROUGE-L score of 0.5835, representing improvements of 52.4% and 15.2% respectively compared to the base BLIP model.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Bootstrapping Language-Image Pre-training (BLIP), Cultural-Aware Adapter, Ekstraksi Fitur, Image Captioning, Ukiran Bali,Cultural-Aware Adapter, Feature Extraction, Image Captioning, Balinese Carving
Subjects:	Q Science Q Science > QA Mathematics Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76 Computer software
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	I Putu Bagus Gede Prasetyo Raharja
Date Deposited:	23 Jan 2026 09:28
Last Modified:	23 Jan 2026 09:28
URI:	http://repository.its.ac.id/id/eprint/130269

Actions (login required)

View Item