Anotace:
This paper presents a detection method of DCAM-YOLOv5 for ground penetrating radar (GPR) to address the difficulty of identifying complex and multi-type defects in tunnel linings. The diversity of tunnel-lining defects and the multiple reflections and scattering caused by water-bearing defects make GPR images quite complex. Although existing methods can identify the position of underground defects from B-scans, their classification accuracy is not high. The DCAM-YOLOv5 adopts YOLOv5 as the baseline model and integrates deformable convolution and convolutional block attention module (CBAM) without adding a large number of parameters to improve the adaptive learning ability for irregular geometric shapes and boundary fuzzy defects. In this study, dielectric constant models of tunnel linings are established based on the electromagnetic simulation software (GPRMAX), including rebar and various structural defects. The simulated and field GPR B-scan images show that the DCAM-YOLOv5 method has better results for detecting different types of defects than other methods, which validates the effectiveness of the proposed detection method.