GPR Subsurface Target ROI Extraction Algorithm And Realization

In this thesis, the theory and algorithm of GPR ROI extraction and the targets detection are investigated. Based on the priori knowledge of a point scattering - hyperbolic feature of subsurface objects in GPR B-Scan images, a method of ROI extraction with Neural Network is presented. According to the characteristics of GPR interpretation, this thesis has presented a method of the B-Scan images pre-processing. The corresponding structure, method of training and testing of the neural network are discussed. Some B-Scan data of plastic-pipe and other objects collected by a GPR set are used to train and test the Neural Network. The results indicate that this method needs a great deal of various training samples and a lot of priori knowledge. Considering about the speed and the accuracy, an automatic detection method of GPR targets is presented in this thesis, which has two steps. First, based on the analysis of GPR A-Scan signal's statistical features, a windowed statistical method is introduced to efficiently extract the ROI from a large amount of data. Then, a modified Hough transform is used to detect the hyperbola from the pre-processed ROI data, which aims to get rid of the false targets and locate the targets accurately. The method had been employed to find and locate targets from GPR data. The results show that all targets are detected correctly, and the depth and horizontal location errors are less than lcm. Moreover, the signal-to-clutter ratio (SCR) of ROI data can fall to about 2dB. At last, the automatic target detection software based on the method discussed above is introduced, and the results of GPR data with this software show its own practicability and validity.