Irsal, Riyandi Banovbi Putera and Prof. Dr.Eng. Fitri Utaminingrum, S.T., M.T. and Kohichi Ogata, Prof. (2024) Improvisasi Model Yolov7 untuk Deteksi Kerusakan Jalan. Magister thesis, Universitas Brawijaya.
Abstract
Hampir setiap negara di dunia menggunakan jalan raya untuk menghubungkan wilayah mereka, dengan total panjang mencapai 64 juta kilometer. Di Indonesia, panjang jalan raya mencapai 550 ribu kilometer, meskipun sebagian besar masih tidak layak dan berbahaya. Deteksi dini kerusakan jalan penting untuk mencegah kerusakan yang lebih parah dan menurunkan biaya pemeliharaan. Metode deteksi kerusakan jalan mencakup inspeksi manual, inspeksi otomatis, dan teknik pemrosesan gambar. Inspeksi manual mahal dan memakan waktu, sementara inspeksi otomatis dengan sensor juga kompleks. Pemrosesan gambar, terutama dengan teknologi deep learning, menawarkan solusi yang lebih efisien dan luas. Penelitian ini bertujuan untuk meningkatkan metode deteksi kerusakan jalan menggunakan dataset Road Damage Detection dengan YOLOv7 yang mengintegrasikan Swin Transformer. Penelitian sebelumnya menunjukkan bahwa YOLO, terutama YOLOv7, efektif untuk deteksi objek real time. YOLOv7 menggunakan tiga metode utama: grid-based detection, bounding box regression, dan Intersection over Union (IoU) untuk menghindari tumpang tindih objek. Penelitian ini juga mengimplementasikan modul CBAM dan SPPFCSPC untuk meningkatkan kinerja deteksi objek. Swin Transformer menggunakan mekanisme self-attention dan strategi Shifted Windows untuk efisiensi komputasi, memberikan hasil yang kompetitif dalam berbagai tugas pengenalan visual. Modul CBAM meningkatkan fokus pada fitur relevan, sedangkan SPPFCSPC dioptimalkan untuk deteksi objek cepat. Hasil penelitian menunjukkan peningkatan signifikan dalam nilai mAP dan efisiensi waktu komputasi. Modifikasi STCSPC pada YOLOv7 meningkatkan kinerja secara signifikan, dengan varian STCSPC, STCSPC2, dan STCSPC3 menunjukkan peningkatan mAP dan F1-score. Penambahan modul CBAM dan SPPF juga meningkatkan kinerja deteksi, dengan varian YOLOv7 CBAM dan SPPF menunjukkan hasil yang baik. Pengujian pada AI Center menunjukkan bahwa model YOLOv7 dan YOLOv7 SPPF sangat efisien dalam pemrosesan, meskipun penambahan modul CBAM dan SPPF cenderung meningkatkan waktu inferensi. Secara keseluruhan, integrasi modul seperti STCSPC, CBAM, dan SPPF pada YOLOv7 menunjukkan potensi besar dalam meningkatkan kinerja deteksi kerusakan jalan.
English Abstract
Almost every country in the world uses highways to connect their territories, with a total length of 64 million kilometers. In Indonesia, the length of highways reaches 550 thousand kilometers, although most of them are still unfit and dangerous. Early detection of road defects is important to prevent further damage and lower maintenance costs. Road defect detection methods include manual inspection, automated inspection, and image processing techniques. Manual inspections are expensive and time-consuming, while automated inspections with sensors are also complex. Image processing, especially with deep learning technology, offers a more efficient and extensive solution. This research aims to improve the road damage detection method using Road Damage Detection dataset with YOLOv7 integrating Swin Transformer. Previous research shows that YOLO, especially YOLOv7, is effective for real-time object detection. YOLOv7 uses three main methods: grid-based detection, bounding box regression, and Intersection over Union (IoU) to avoid object overlap. This research also implements CBAM and SPPFCSPC modules to improve object detection performance. Swin Transformer utilizes self-attention mechanism and Shifted Windows strategy for computational efficiency, providing competitive results in various visual recognition tasks. The CBAM module enhances the focus on relevant features, while SPPFCSPC is optimized for fast object detection. The results show significant improvements in mAP values and computational time efficiency. The STCSPC modifications to YOLOv7 improved performance significantly, with STCSPC, STCSPC2, and STCSPC3 variants showing improved mAP and F1-score. The addition of CBAM and SPPF modules also improved detection performance, with YOLOv7 CBAM and SPPF variants showing good results. Tests on the AI Center show that the YOLOv7 and YOLOv7 SPPF models are highly efficient in processing, although the addition of CBAM and SPPF modules tends to increase inference time. Overall, the integration of modules such as STCSPC, CBAM, and SPPF in YOLOv7 shows great potential in improving the performance of road defect detection.
| Item Type: | Thesis (Magister) |
|---|---|
| Identification Number: | 042415 |
| Uncontrolled Keywords: | Yolo, Swin Transformer, Attention Based, Road Damage Detection |
| Divisions: | S2/S3 > Magister Ilmu Komputer, Fakultas Ilmu Komputer |
| Depositing User: | Sugeng Moelyono |
| Date Deposited: | 31 Oct 2024 06:46 |
| Last Modified: | 31 Oct 2024 06:46 |
| URI: | http://repository.ub.ac.id/id/eprint/227979 |
![[thumbnail of v]](http://repository.ub.ac.id/style/images/fileicons/text.png) |
Text (v)
Riyandi Banovbi Putera Irsal.pdf Restricted to Registered users only Download (7MB) |
Actions (login required)
 |
View Item |