Automatic Recognition of Hidden Road Defects from GPR Images Using an Enhanced CNN Approach

Abstract

Ground-penetrating radar (GPR) has been widely used to nondestructively detect hidden defects in road structures. However, the interpretation of GPR images relies on the subjective experience of technicians, which is demanding, inefficient, and leads to the misinterpretations of targeted objects. This paper establishes a new framework. First, data enhancement (e.g., mirroring, random noise, etc.) was used to elevate the generalization of the proposed algorithm. Second, the receptive-field attention (RFA) and squeeze-and-excitation network have been integrated into you only look once (YOLO), resulting in RFS-YOLO, for intelligent recognition of hidden road defects. Finally, adding gradient-weighted class activation mapping further enhanced the interpretability of the model and the precision of small target detection. RFS-YOLO employs RFAConv to replace standard convolutional layers, significantly boosting network performance without a substantial increase in parameter or computational costs. Further, incorporation of an squeeze-and-excitation (SE) attention module allows the model to concentrate more intently on the features of the defects of interest. The study conducted an ablation experiment to demonstrate the explicit details of each modified part of the improved framework. In comparison with the traditional deep learning model, the proposed algorithm owns higher predicting precision (93.2%) and mean average precision (95.3%), which suggests that the developed algorithm has better detection performance and robustness. This work casts new light on the applications of GPR in the automatic recognition of hidden defects.