Research On Classification And Detection Method Of Ground Penetrating Radar Underground Target

Ground Penetrating Radar(GPR)is a device for remote sensing of underground scenes using electromagnetic pulses,which is widely used in various engineering fields with the advantages of non-destructive detection and high detection efficiency.Therefore,intelligent interpretation of ground penetrating radar data has become a hot research topic in the field.The task of underground target classification and detection is the key to the intelligent interpretation of ground penetrating radar data,which is of great significance to the analysis of underground target properties and the construction of underground scenes.In this paper,focusing on the pain point of complicated ground penetrating radar data and the low efficiency of the traditional way to classify and identify underground targets,we explore the automated classification and detection method of underground targets based on the convolutional neural network algorithm.The main work of this paper is as follows:1.To address the shortcomings of ideal targets and simple scenes in ground penetrating radar data obtained by using electromagnetic simulation software,according to the working principle and detection process of ground penetrating radar system,a large number of detection experiments are conducted using the group’s 3D array ground penetrating radar system to obtain the real data.By analyzing the feature performance of underground targets in ground penetrating radar B-scan data,the real B-scan image data are screened and clipped for targets.Finally,based on the problems and data characteristics in the analysis results of the real data,we propose solutions and construct the real dataset of underground targets.2.To address the problem of scarcity of effective data containing underground targets in ground penetrating radar real data,an underground target detection method with small sample data set is proposed,and a small sample data set of underground targets is constructed.The method designs an improved U-Net target segmentation network under attention constraint,which can segment targets efficiently and accurately despite the shortage of training data;based on target segmentation map,target detection is performed by calculating the connected region and designing a target feature key point localization algorithm to locate target key points in the detection box.The experimental results show that the segmented images predicted by this method are more accurate than those predicted by the basic U-Net network,and the Dice similarity coefficient is improved by 13%,which guarantees the accuracy of target detection and key point localization and verifies the superiority of this network.3.To address the problem that the mainstream target detection algorithms need to manually label the target anchor boxes and there are many interfering factors in the labeling process of ground penetrating radar real B-scan images,we study a underground target classification and detection method that can automatically detect the target features in B-scan data without pre-labeling anchor boxes.The result of image classification is obtained by designing a residual classification network with the attention mechanism,then calculating the class-discriminative localization map based on the classification result,and finally segmenting the class-discriminative localization map based on the HSV color model to obtain the feature area of the target.The experiments show that compared with Res Net50,the classification network in this paper can focus on the target feature information and shield the background noise,the target feature areas in the class-discriminative localization map are significant,and the target detection accuracy is high,the average center distance between the recognition frame and the real frame decreases by 53.1%,and the average IOU increases by 49.7%,which verifies the effectiveness of the method in this paper.