Integrasi YOLOv8 Dengan Slicing Aided Hyper Inference (SAHI) Untuk Deteksi Karakter Aksara Pegon

Rahmah, Rajni Yafi' Amelia (2026) Integrasi YOLOv8 Dengan Slicing Aided Hyper Inference (SAHI) Untuk Deteksi Karakter Aksara Pegon. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5002221009-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (1MB) | Request a copy

Abstract

Transliterasi aksara Pegon ke aksara Latin dibutuhkan untuk memastikan teks dapat dipahami oleh masyarakat. Saat ini, terdapat 400 naskah pegon yang terdaftar oleh Perpustakaan Nasional dan masih banyak naskah lainnya yang tidak terdaftar. Namun proses transliterasi saat ini masih didominasi dengan pendekatan manual yang membutuhkan waktu cukup lama. Computer Vision dengan metode You Only Look Once (YOLO) dapat digunakan untuk proses transliterasi. Namun YOLO masih memiliki keterbatasan dalam mendeteksi karakter aksara Pegon yang rapat dan berhimpit dalam satu naskah utuh. Pada penelitian Tugas Akhir ini penggunaan YOLO diintegrasikan dengan framework Slicing Aided Hyper Interface (SAHI). SAHI digunakan sebagai alat pembantu YOLO untuk mengenali objek kecil dan berhimpit yang berbasis slicing. Proses deteksi YOLO-SAHI diawali dengan citra utuh sebagai inputan yang selanjutnya dikenai proses slicing dengan ukuran 640 x 640, tiap citra hasil slicing dideteksi dengan model deteksi YOLOv8 yang telah di fine-tuning sehingga menghasilkan deteksi pada tiap aksara Pegon. Hasil deteksi pada tiap citra slicing selanjutnya digabungkan kembali menjadi citra utuh dan dikenai NMS untuk menghilangkan deteksi yang bertumpuk. Hasil integrasi YOLO-SAHI berhasil menangkap detail pada objek aksara yang padat dan berhimpit dengan nilai recall 0.8216, mAP@50 0.8103, dan presentase keberhasilan mendeteksi kelas sebesar 96.84%
=================================================================================================================================
Transliterating Pegon script into Latin script is essential to ensure the text is comprehensible to the public. Currently, the National Library has registered 400 Pegon
manuscripts, yet many others remain unregistered. However, the prevailing transliteration process relies on manual methods that are time-consuming. Computer Vision utilizing the ”You Only Look Once” (YOLO) method offers a potential solution for this process; nevertheless, YOLO faces limitations in detecting Pegon characters that are densely packed and overlapping within a complete manuscript. In this final project research, YOLO is integrated with the Slicing Aided Hyper Interface (SAHI) framework. SAHI serves as a slicing-based aid to assist YOLO in recognizing small and overlapping objects. The YOLO-SAHI detection process begins with the full image as input, which is then subjected to a slicing process using 640 x 640 dimensions; each sliced image is processed by a fine-tuned YOLOv8 detection model to identify individual Pegon characters. The detection results from each slice are subsequently reassembled into the
full image, and Non-Maximum Suppression (NMS) is applied to eliminate overlapping detections. The YOLO-SAHI integration successfully captures details of dense and overlapping character objects, achieving a recall of 0.8216, an mAP@50 of 0.8103, and a class detection success rate of 96.84

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Aksara Pegon, Slicing Aided Hyper Interface, Transliterasi, You Only Look Once, PegonScript, Slicing Aided Hyper Interface, Transliteration, You Only Look Once
Subjects:	L Education > L Education (General) P Language and Literature > PI Oriental languages and literatures Q Science > QA Mathematics Q Science > QA Mathematics > QA76.6 Computer programming. Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Mathematics > 44201-(S1) Undergraduate Thesis
Depositing User:	Rajni Yafi' Amelia Rahmah
Date Deposited:	28 Jul 2026 03:16
Last Modified:	28 Jul 2026 03:16
URI:	http://repository.its.ac.id/id/eprint/138428

Actions (login required)

View Item