Effendy, Moch. Rafy Adhipramana (2026) Pengembangan Arsitektur YOLOv13n Melalui Penambahan Detection Head Dan Modifikasi Konvolusi Untuk Peningkatan Performa Deteksi Objek Kecil. Masters thesis, Institut Teknologi Sepuluh Nopember.
![[thumbnail of 6026242012-Master_Thesis.pdf]](http://repository.its.ac.id/style/images/fileicons/text.png) |
Text
6026242012-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (10MB) | Request a copy |
Abstract
Penelitian ini mengusulkan modifikasi arsitektur YOLOv13n untuk meningkatkan performa deteksi objek kecil dengan tetap menjaga efisiensi komputasi. Modifikasi yang diterapkan meliputi penambahan Head P2, pengurangan Stride-1, penggunaan Kernel 5×5, dan penerapan Partial Convolution (PConv). Keempat modifikasi diuji secara individual maupun kombinasi pada dataset Dolphin14k, TT100K, dan VisDrone2019. Hasil pengujian menunjukkan bahwa kombinasi seluruh modifikasi (Combined) berhasil meningkatkan mAP@50 melebihi target minimal 5% dibandingkan Baseline, yaitu dari 0,6116 menjadi 0,6770 (naik 6,54%) pada Dolphin14k, dari 0,6022 menjadi 0,7504 (naik 14,82%) pada TT100K, dan dari 0,2906 menjadi 0,3810 (naik 9,04%) pada VisDrone2019. Modifikasi Stride-1 dan Head P2 secara konsisten memberikan peningkatan akurasi terbesar, sedangkan PConv menjadi modifikasi paling efisien dari sisi beban komputasi dan waktu inferensi, meskipun pengaruhnya terhadap akurasi bergantung pada karakteristik dataset. Hasil penelitian ini menunjukkan adanya trade-off antara akurasi deteksi dan efisiensi komputasi, sehingga pemilihan kombinasi modifikasi perlu disesuaikan dengan kebutuhan aplikasi dan karakteristik dataset.
==========================================================================================================================================
This research proposes architectural modifications to the YOLOv13n model to improve small object detection performance without excessively compromising computational efficiency. The proposed modifications include adding a detection head at the P2 level to enrich high-resolution feature coverage, reducing the convolution stride (Stride-1) to preserve spatial information, enlarging the convolution kernel to 5×5 (Kernel-5) to expand the receptive field, and applying Partial Convolution (PConv) to reduce computational load and accelerate inference. The four modifications were evaluated individually and in combination on three small-object datasets, namely Dolphin14k, TT100K, and VisDrone2019. The results show that combining all modifications (Combined) increased mAP@50 over the Baseline beyond the minimum 5% target, from 0.6116 to 0.6770 (a 6.54% increase) on Dolphin14k, from 0.6022 to 0.7504 (a 14.82% increase) on TT100K, and from 0.2906 to 0.3810 (a 9.04% increase) on VisDrone2019. Stride-1 and Head P2 consistently contributed the largest accuracy gains, while Partial Convolution was the most computationally efficient modification in terms of both computational load and inference time, although its effect on accuracy was dataset-dependent. These findings confirm a trade-off between detection accuracy and computational efficiency, indicating that the optimal combination of modifications should be tailored to application requirements and dataset characteristics.
| Item Type: | Thesis (Masters) |
|---|---|
| Uncontrolled Keywords: | YOLOv13n, Deteksi Objek Kecil, Partial Convolution, Receptive Field, mAP ========================================================================================================================YOLOv13n, Small Object Detection, Partial Convolution, Receptive Field, mAP |
| Subjects: | T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques. Image analysis--Data processing. |
| Divisions: | Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Information System > 59101-(S2) Master Thesis |
| Depositing User: | Moch. Rafy Adhipramana Effendy |
| Date Deposited: | 30 Jul 2026 06:13 |
| Last Modified: | 30 Jul 2026 06:13 |
| URI: | http://repository.its.ac.id/id/eprint/139128 |
Actions (login required)
 |
View Item |