A Lightweight Detection Algorithm for Pavement Structure Distress Based on Ground-Penetrating Radar Investigation and Joint Pruning

Edge computing technology enables in situ diagnosis based on neural networks in ground-penetrating radar (GPR) detection. However, there is a contradiction between the limited resources of radar terminal and the high resource demands of the pavement distress detection model. This paper proposed a lightweight GPR detection model based on joint pruning to address this issue. First, the model uses YOLOv5s as the basic model and EfficientNet as the backbone feature extraction network in training, achieving structure compression of the feature extraction network. Next, the filter pruning via geometric median is used to further remove the redundant filters of each convolutional layer. In addition, considering the distress diversity and background complexity, attention mechanisms are incorporated to enhance the algorithm’s robustness. A dataset of structural distresses, including loose, cracked, empty, and poor interlayer, was constructed and brought into the training. The results show improvements in both identification accuracy and detection speed, with the mean average precision (mAP) reaching 91.5% and giga floating-point operations per second (Gflops) decreasing by 21.74%. The proposed method can automatically and efficiently extract information from GPR images, thus providing technical support for the deployment of radar equipment on resource-limited edge devices.