Kombinasi Model YOLOv8 dan EfficientNet-B0 untuk Identifikasi Karakter Aksara Pegon

Tsurayyaa, Qayla Syadza (2026) Kombinasi Model YOLOv8 dan EfficientNet-B0 untuk Identifikasi Karakter Aksara Pegon. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5002221171-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (1MB) | Request a copy

Abstract

Aksara Pegon merupakan sistem tulisan berbasis Arab yang digunakan dalam berbagai dokumen keagamaan dan sastra di Indonesia. Kemiripan bentuk antarkarakter, penggunaan harakat dan diakritik, serta keterbatasan data digital menjadi kendala dalam proses identifikasi karakter aksara Pegon secara otomatis. Penelitian ini bertujuan menerapkan dan mengevaluasi sistem identifikasi karakter aksara Pegon ketik melalui pipeline dua tahap yang menggabungkan YOLOv8 untuk mendeteksi lokasi karakter dan EfficientNet-B0 untuk mengklasifikasikan jenis karakter. Dataset yang digunakan terdiri atas 27 citra halaman kitab Aqidatul Awam yang mencakup 95 kelas karakter dan dibagi berdasarkan halaman menjadi data pelatihan, validasi, dan pengujian dengan rasio 70:15:15. Model YOLOv8 dilatih menggunakan dua skenario, yaitu deteksi satu kelas objek dan deteksi 95 kelas karakter, sedangkan EfficientNet-B0 dilatih menggunakan pendekatan transfer learning dan strategi two-stage fine-tuning. Hasil pengujian menunjukkan bahwa YOLOv8 dengan satu kelas objek menghasilkan performa terbaik dengan nilai precision sebesar 0,9397, recall sebesar 0,9079, mAP50 sebesar 0,9295, dan mAP50–95 sebesar 0,5915. Model EfficientNet-B0 menghasilkan Top-1 Accuracy sebesar 93,16%, Top-5 Accuracy sebesar 98,31%, dan F1-score macro sebesar 87,77% pada data pengujian klasifikasi. Pada pengujian sistem gabungan, sebanyak 962 dari 1.030 karakter ground truth berhasil dicocokkan dengan hasil deteksi, dan 860 karakter berhasil diklasifikasikan dengan benar. Hasil tersebut menghasilkan akurasi klasifikasi pada potongan citra yang cocok sebesar 89,40% dan akurasi keseluruhan sistem secara end-to-end sebesar 83,50%. Berdasarkan hasil tersebut, kombinasi YOLOv8 dan EfficientNet-B0 dapat diterapkan untuk mengidentifikasi karakter aksara Pegon untuk mendukung percepatan pelestarian dan digitalisasi dokumen historis Islam Nusantara.
=========================================================================================================================================
Pegon script is an Arabic-based writing system used in various religious and literary documents in Indonesia. Similarities among character shapes, the presence of vowel marks and diacritics, and the limited availability of digital data create challenges for automatic Pegon character identification. This study aims to implement and evaluate a two-stage pipeline for identifying typed Pegon characters by combining YOLOv8 for character localization and EfficientNet-B0 for character classification. The dataset consists of 27 scanned pages from the Aqidatul Awam book, covering 95 character classes, and was divided by page into training, validation, and test sets using a ratio of 70:15:15. The YOLOv8 model was trained using two scenarios, namely single-class object detection and 95-class character detection, while EfficientNet-B0 was trained using transfer learning and a two-stage fine-tuning strategy. The test results show that the single-class YOLOv8 model achieved the best detection performance, with a precision of 0.9397, a recall of 0.9079, an mAP50 of 0.9295, and an mAP50–95 of 0.5915. The EfficientNet-B0 model achieved a Top-1 Accuracy of 93.16%, a Top-5 Accuracy of 98.31%, and a macro F1-score of 87.77% on the classification test data. In the combined pipeline evaluation, 962 out of 1,030 ground-truth characters were successfully matched with the YOLOv8 detections, and 860 characters were correctly classified. These results produced a classification accuracy of 89.40% for the matched character crops and a full end-to-end accuracy of 83.50%. The results indicate that the combination of YOLOv8 and EfficientNet-B0 can be applied to Pegon character identification, although improvements in the detection and character-cropping stages are still required to enhance the overall system performance.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Aksara Pegon, OCR, YOLOv8, EfficientNet-B0, Pegon Script, OCR, YOLOv8, EfficientNet-B0
Subjects:	T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques. Image analysis--Data processing.
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Mathematics > 44201-(S1) Undergraduate Thesis
Depositing User:	Qayla Syadza Tsurayyaa
Date Deposited:	29 Jul 2026 06:11
Last Modified:	29 Jul 2026 06:11
URI:	http://repository.its.ac.id/id/eprint/139327

Actions (login required)

View Item