Deep learning-enhanced numerical simulation of ground penetrating radar and image detection of road cracks
Aiming at the requirements of nondestructive testing and intelligent recognition of cracks inside the pavement structures,a ground penetrating radar(GPR)detection and automatic identification method based on improved you only look once version 8(YOLOv8)and AutoAugment enhancement was proposed.Combined with Gprmax numerical simulation,laboratory model test,and field GPR measurement,it is concluded that the internal cracks showed a prominent"hyperbola"feature on GPR images,and the width of the cracks is proportional to the size of the hyperbola.MALA GPR with GX-750 was used to detect the crack images inside the pavement structure,and a 640X640 pixel frame was used to intercept the image after multiple filtering processing.Aiming at the problem of small crack features in GPR images,the improved YOLOv8 model was obtained by adding an output layer with pixels of 160 × 160 to the latest YOLOv8 model.Meanwhile,Selective Kernel Networks(SKNet)attention mechanism was introduced to increase the sensitivity field further and adopted the loss function of power IoU to decrease the training loss of the model.As for the original image data set,the AutoAugment method with unsupervised automatic augment was applied to find the best enhancement strategy and its probability and intensity through the reinforcement learning algorithm optimized by the proximal strategy to realize the effective expansion of the GPR image data set.After training and testing on the expanded data set,the results show that the improved YOLOv8 model has achieved a mean average precision(mAP)of 90.7％and an Fl score of 90.1％,which is 6.3％and 5.9％higher than that of the original YOLOv8 model,and also significantly exceeds the mainstream target detection model.After image enhancement,the mAP and F1 score of the model increased by 4.1％and 4.6％,respectively,showing good robustness to crack image detection of various scales.The application of the GPR image intelligent recognition method in the maintenance section is consistent with the coring verification results,which shows that the improved model is reliable in practical engineering applications.
NETL
NSTL
万方数据
