Research On GPR Image Target Region Extraction And Material Classification Algorithm Based On YOCAMR And RESVT

GPR is a non-destructive detection method based on electromagnetic waves,which is widely used in municipal engineering,military rescue and other fields.With the continuous improvement of people’s demand for underground area detection,higher requirements are also put forward for the signal processing technology of GPR.The key issues include rapid detection and framing of objects in massive data and material classification and identification of underground objects.This thesis studies the GPR target region extraction and material classification methods based on deep learning technology.This thesis first briefly expounds the detection principle of GPR,and analyzes the research status and development trend of deep learning technology applied to GPR image processing.Aiming at the problem of low accuracy of target area framing in GPR B-scan images,the interpretability of deep learning model is innovatively applied to GPR target area extraction,and the YOCAMR target area detection algorithm is proposed.Aiming at the problem of GPR’s classification of underground target material properties,a solution method based on Transformer framework is studied,and a RESVT model framework is proposed.The simulation and measured data sets are produced for network training and testing,which verifies the effectiveness of the proposed algorithm.The specific work is as follows:(1)This thesis proposes a YOCAMR algorithm for target region detection based on GPR images.The algorithm first extracts hyperbolic regions based on the YOLOv5 framework.Then,deep learning interpretability is introduced into the target area extraction of GPR images.With the help of the model training process,the GPR images are presented in the form of CAM heat maps,and the pixel areas with rich information are extracted by dynamically adjusting the threshold.Finally,the two regions are merged.This algorithm makes up for the inaccuracy of the traditional single model relying on the manual labeling process.At the same time,from the perspective of deep learning interpretability,it considers the physical information contained in the B-scan image itself to achieve accurate pixel-level region extraction.more applicable.Based on the simulation data and the measured data,experimental tests were carried out respectively,and the Res Net50 network was used to quantitatively evaluate the three different extraction regions to verify the effectiveness of the YOCAMR algorithm.(2)This thesis proposes a RESVT framework for the GPR material classification problem.This thesis studies the Vi T model based on the Transformer framework,adjusts the network skeleton on the basis of analyzing the original network structure,and proposes the RESVT model framework based on the feature fusion of CNN and Transformer.The model takes into account the advantages of the self-attention mechanism and the convolutional neural network model.The improved RESVT model structure is used for the GPR material classification experiment,and the accuracy of the measured GPR image material classification experiment reaches 98.56%.At the same time,compared with other network models,the RESVT model has higher accuracy in the experiment,which verifies the effectiveness of the model.(3)Data set acquisition and comparison and analysis of experimental results.In this thesis,simulation and measured data sets of different material categories are obtained through Gpr Max software and laboratory SIR4000 equipment.At the same time,a total of five groups of experiments were set up to compare the accuracy rates to verify the effectiveness of the above algorithms and models.Compared with the traditional YOLOv5 and Res Net50 models,the YOCAMR algorithm proposed in this thesis improves the accuracy of the measured data set by 6.58% and 1.08%,respectively.Similarly,the image classification accuracy of the RESVT model is 5.39% and 5.75% higher than that of the Vi T model and the Res Net50 model,respectively,in the measured dataset.It confirms the practical significance of the algorithm proposed in this thesis in the actual scene of the GPR material classification problem.