Road Disease Detection Based on Improved YOLOv8s Algorithm

With the rapid development of China's transportation infrastructure and the continuous expansion of the highway network, the highway system is gradually transitioning from the construction phase to the maintenance and management phase. A variety of road disease detection methods have emerged. T...

Full description

Saved in:
Bibliographic Details
Published in2024 6th International Conference on Data-driven Optimization of Complex Systems (DOCS) pp. 594 - 600
Main Authors Wu, Rile, Sun, Xuelian
Format Conference Proceeding
LanguageEnglish
Published IEEE 16.08.2024
Subjects
Attention mechanisms
Deformable Attention Mechanism
Diseases
GD-Mechanism
Information exchange
Maintenance
MPDIoU
Neck
Nonhomogeneous media
Optimization
Road Disease
Roads
Transportation
YOLO
YOLOv8
Online AccessGet full text

Cover

More Information
Summary:With the rapid development of China's transportation infrastructure and the continuous expansion of the highway network, the highway system is gradually transitioning from the construction phase to the maintenance and management phase. A variety of road disease detection methods have emerged. This paper focuses on road disease as the research subject and conducts a study on road disease detection based on computer vision, proposing a detection model suitable for modern highways. Firstly, the attention mechanism Dattention, which has deformable attention mechanism and dynamic sampling points, is first introduced and added to the C2f unit block of YOLOv8. Dattention introduces a deformable attention mechanism, which focuses only on a small critical area of the image. This method significantly lowers computational requirements while maintaining excellent performance. Secondly, a GD network with aggregation-distribution mechanism is designed in the Neck part of YOLOv8 for effective information exchange in YOLO. The ability of the Neck structure in information fusion is significantly enhanced by globally fusing multilayer features and introducing higher-level global information, improving the performance of the model on different object sizes. Using MPDIoU instead of CIoU improves the positioning precision and categorization performances of the detected frames. Experimental findings indicate that the improved method achieves a 70.5% mAP on the RDD_Japan dataset, marking a 2% increase over the original model.
DOI:10.1109/DOCS63458.2024.10704316