Pengenalan Berbasis Keterkaitan Pola Dan Deep Learning Untuk Temu Kembali Citra Ukiran Bali

Darma, I Wayan Agus Surya (2022) Pengenalan Berbasis Keterkaitan Pola Dan Deep Learning Untuk Temu Kembali Citra Ukiran Bali. Doctoral thesis, Institut Teknologi Sepuluh Nopember.

Text 05111960010013-Doctoral.pdf Restricted to Repository staff only Download (20MB)

Abstract

Ukiran Bali adalah hiasan ukiran pada bangunan di Bali dengan ragam hias motif khas Bali. Motif ukiran Bali terdiri dari beberapa jenis pola yang mewakili nilai-nilai masyarakat di Bali. Warisan budaya ini dapat dilestarikan dengan mengumpulkan data digital ukiran Bali yang tersebar pada bangunan pura di Bali. Sistem temu kembali dapat digunakan untuk mengelola data digital ukiran Bali dan memudahkan mendapatkan informasi terkait motif ukiran Bali. Penelitian ini bertujuan untuk membangun metode pengenalan motif ukiran Bali yang dapat digunakan pada proses temu kembali ukiran Bali. Tantangan dalam pengenalan motif ukiran Bali yaitu: (1) terbatasnya jumlah data dan belum tersedianya dataset ukiran Bali, (2) keunikan motif ukiran Bali yang memiliki variasi yang tinggi, dan (3) Ukiran Bali yang disusun atas sub-motif ukiran yang saling terkait. Penelitian ini bertujuan untuk membangun metode pengenalan motif ukiran Bali berbasis keterkaitan pola yang terdiri dari model augmentasi data, model deteksi dan segmentasi, dan model pengenalan untuk temu kembali ukiran Bali. Pertama, model augmentasi data menerapkan neural style transfer (NST) dan geometric transformation (GT) untuk menghasilkan citra sintetis baru dengan variasi citra berdasarkan warna dan posisi geometri citra yang mewakili berbagai kondisi pengambilan citra ukiran Bali. Berdasarkan hasil augmentasi data dan proses anotasi, dibangun dataset baru ukiran Bali untuk tugas klasifikasi, deteksi dan segmentasi motif ukiran Bali. Kedua, model deteksi sub-motif ukiran Bali dikembangkan berbasis You Only Look Once (YOLO) dengan menerapkan network scaling dan ensemble model berbasis non-maximum suppression (NMS). Model segmentasi dibangun berbasis Mask R-CNN untuk melakukan tugas segmentasi motif ukiran Bali. Ketiga, model pengenalan dibangun berdasarkan karakteristik antar sub-motif ukiran Bali yang saling terkait. Model pengenalan dibangun berbasis graph convolutional network (GCN) dan convolutional neural network (CNN) dengan mengkombinasikan fitur graf dengan fitur citra ukiran Bali untuk meningkatkan kinerja pengenalan. Model pengenalan yang dihasilkan pada proses ini, selanjutnya digunakan pada proses temu kembali ukiran Bali. Berdasarkan hasil penelitian, pendekatan temu kembali menggunakan pengenalan motif berbasis keterkaitan pola menghasilkan kinerja temu kembali vi yang dapat mengungguli varian baseline ResNet. Selain itu, mengungguli deskriptor fitur konvensional seperti Histogram of Oriented Gradient (HoG) dan Hue, Saturation, Value (HSV). Setiap tahapan pada metode yang diusulkan mendukung hasil ini. Pertama, model augmentasi NST dan GT digunakan untuk membangun dataset ukiran Bali yang terdiri dari 2.364 citra dan 9.326 label. Dataset ini diujikan menggunakan MobileNet, Inception-v3, VGG16, dan VGG19 menghasilkan kinerja klasifikasi citra motif ukiran Bali hingga 91,6%. Kedua, model deteksi sub-motif penyusun ukiran Bali berbasis network scaling dan NMS ensemble menghasilkan kinerja deteksi mencapai 98% dengan mengurangi waktu pelatihan dan ukuran parameter masing-masing sebesar 62% dan 84,4%. Ketiga, model pengenalan berbasis keterkaitan pola dapat menghasilkan kinerja dengan akurasi 98,93% yang mengungguli model-model pembanding.
==============================================================================================================================
Balinese carvings are decorations made by carving buildings with various Balinese motifs. Balinese carving motifs consist of several types of patterns that represent the values of the people in Bali. This cultural heritage can be preserved by collecting digital data on Balinese carvings scattered on temple buildings in Bali. The retrieval system can be used to manage digital data for Balinese carvings and make it easier to get information related to Balinese carving motifs. This study aims to develop a method for recognizing Balinese carving motifs that can be used in retrieving Balinese carvings. The challenges in recognizing the Balinese carving motifs are: (1) the limited amount of data and the unavailability of a Balinese carving dataset, (2) the uniqueness of Balinese carving motifs which have a high variation, and (3) Balinese carvings which are composed of interrelated carving sub-motifs. This study aims to build a method of recognizing Balinese carving motifs based on pattern relationship consisting of data augmentation models, detection and segmentation models, and recognition models for Balinese carving retrieval. First, the data augmentation model applied neural style transfer (NST) and geometric transformation (GT) to generate a new synthetic image with image variations based on color and image geometry positions representing various conditions for taking Balinese carvings. Based on data augmentation and annotation processes, a new dataset of Balinese carvings was built for classifying, detecting, and segmenting Balinese carving motifs. Second, the carving sub motifs detection model was developed based on You Only Look Once (YOLO) by applied network scaling and ensemble models based on non-maximum suppression (NMS). The segmentation model was built based on Mask R-CNN to perform the task of segmenting Balinese carving motifs. Third, the recognition model was built based on the interrelated characteristics of Balinese carving sub-motifs. The recognition model was built based on a graph convolutional network (GCN) and a convolutional neural network (CNN) by combining graph features with Balinese carving image features to viii improve recognition performance. Then, the recognition model is then used to retrieve Balinese carvings. Based on the research results, the image retrieval approach based on pattern-relationship recognition, produced retrieval performance that can outperform the baseline ResNet variant. In addition, the proposed model outperformed conventional feature descriptors, i.e., Histogram of Oriented Gradient (HoG) and Hue, Saturation, Value (HSV). Each step in the proposed method supported this result. First, the NST and GT augmentation models were used to build a Balinese carving dataset consisting of 2.364 images and 9.326 labels. This dataset was tested using MobileNet, Inception-v3, VGG16, and VGG19 resulting in the classification performance of Balinese carving motifs up to 91,6%. Second, the detection model for Balinese carving sub-motifs based on network scaling dan NMS ensemble resulted in detection performance of up to 98% by reducing the training time and parameter sizes by 62% and 84,4%, respectively. Third, the pattern-relationship-based recognition model (GFF-Carving) could improve model performance with an accuracy of 98.93% which outperformed the comparison models.

Item Type:	Thesis (Doctoral)
Additional Information:	RDIf 006.42 Dar p-1 2022
Uncontrolled Keywords:	augmentasi data, keterkaitan pola, network scaling, neural style transfer, temu Kembali. data augmentation, image retrieval, network scaling, neural style transfer, pattern relationship.
Subjects:	T Technology > T Technology (General) > T58.5 Information technology. IT--Auditing
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55001-(S3) PhD Thesis (Comp Science)
Depositing User:	Mr. Marsudiyana -
Date Deposited:	09 Jul 2026 07:18
Last Modified:	09 Jul 2026 07:18
URI:	http://repository.its.ac.id/id/eprint/134600

Actions (login required)

View Item