Novel YOLOv3 Model With Structure and Hyperparameter Optimization for Detection of Pavement Concealed Cracks in GPR Images

Accurately identifying and localizing concealed cracks in asphalt pavements through nondestructive ground-penetrating radar (GPR) testing has attracted much attention. However, various realistic factors restrict its detection accuracy. Therefore, this paper proposed a novel YOLOv3 model with a ResNet50vd-deformable convolution (DCN) backbone and a hyperparameter optimization (HPO) method using Bayesian search. First, a 3D-Radar system with multi-channel DXG $^{\mathrm {TM}}$ ground-coupled antenna arrays was used to investigate concealed cracks in the asphalt pavement to establish a crack distress dataset with 366 images and 533 cracks. Then, owing to the small GPR image dataset, a simple semi-supervised label distillation (SSLD) method was employed to obtain the pretrained model. Subsequently, Bayesian searching based HPO was performed to find the maximum mean average precision (mAP) and corresponding hyperparameters in 20 searches. Finally, several mainstream detection models were used for comparisons. Experimental results showed that YOLOv3-ResNet50vd-DCN model converged faster and had a smaller loss value (approximately 0.05) in the training process, which illustrates the advantage of model pretraining with the SSLD method. Besides, the proposed model also achieved good detection results, with 92.6% accuracy, 92.3% F1 score, 92.1% mAP, and 0.923 area under the ROC curve (AUC) value, all 3% to 7% higher than that of the other models. After Bayesian searching for HPO, the detection results were further improved to 94.8% accuracy, 96.2% F1 score, 94.6% mAP, and 0.962 AUC value, validating the reliability and superiority of the proposed model and optimization method in detecting pavement concealed cracks of GPR images.