Analysis Of 3D Ground Penetrating Radar Road Viod Based On Convo-lution Neural Network

In response to the social problem of urban roads with potential problems of hollows leading to pavement collapse,ground-penetrating radar detection and deep-learning image recognition techniques are used to research such diseases.The road is scanned by 3D ground-penetrating radar to screen the disease and study its data images;the road deglaciation disease images are simulated to extract different types of disease characteristics,while data enhancement is carried out to establish a database;the classical model is used for training recognition and classification of the disease dataset using the migration learning method,and the recognition effect is compared to select the model with the best performance for improvement and optimization to improve recognition accuracy,and establish the road disease collection and analysis system.According to the electromagnetic wave principle,the radar detection process research and analysis,the use of three-dimensional ground-penetrating radar for deep scanning of heavy traffic sections,through the acquisition of the whole section of the radar image,analysis of the road due to the roadbed structure and geological conditions and other conditions caused by local debonding and collapse and other disease image information,scanning results confirm that three-dimensional ground-penetrating radar is suitable for urban road detection.For the problem of lack of high-quality radar B-scan in the deep learning model,the simulation software Gpr Max is used to simulate the propagation process of electromagnetic waves in the subsurface,generating four types of disease models and B-scan images of the subsurface cavity,cavity-filling,roadbed hollows,and roadbed containing water,and the simulation effect matches with the field data,combined with the actual collected images for expanding the radar data set.The radar images were optimized by considering various factors such as under pavement disease pattern,ground-penetrating radar main frequency,and noise interference,including the addition of Gaussian noise with three signal-to-noise ratios and the processing of disease feature labels,high quality image datasets were obtained for neural network recognition experiments.Using the learning method of model migration,four classical convolutional models,Yolo v5,Yolo v4,SSD,and Faster RCNN,were used to train the recognition of various types of disease images and the training effects were evaluated by accuracy,recall,accuracy,AP index,and F1 score,and finally,the Yolo v5 model with the optimal parameter combination suitable for radar orthorectification was obtained after comparison,which is suitable for continue to improve to enhance the comprehensive effect of recognition.The efficient channel attention mechanism is introduced to purify the feature information;the loss function is improved to expand the target frame regression range,which achieves the purpose of increasing stability and improving the convergence accuracy and other indicators.The prediction effect of the improved Yolo v5 model is evaluated by combining the 3D ground-penetrating radar field disease scan data.The iterative process is relatively smooth,the recognition process expands the target extraction range,improves the image utilization rate,the recognition accuracy and various indexes are improved,and the accurate recognition of radar disease images is achieved.As a whole,the integrated road detection system of 3D ground-penetrating radar and convolutional neural network combines the advantages of fast recognition speed,high detection rate and high recognition accuracy.