Image Processing Of GPR And Its Application In The Detection Of Underground Void

The void is one of the most serious diseases under the road. In recent years, the phenomenon of collapse has happened in many big cities in China because of the void,which serious threat to our life and property safety. It is significant if we detect the void and remedy it timely. In order to accurate and real time identification the void, we use image processing of Ground penetrating radar(GPR) to detect the void.This paper studied several key problems of the above questions, included clutter reduction of image, image’s target positioning and classification the signals of different material.In view of the GPR image clutter reduction, we proposed the clutter suppression method based on PCA combining with gradient magnitude. The results of simulation and experiment show that the proposed algorithm is able to significantly suppress the clutters and is superior to some common methods, like mean subtraction method 、PCA clutter suppression and gradient magnitude clutter suppression.For the positioning of target in B-scan image, we propose an automatic target orientation. We use curve symmetry to compute the horizontal coordinates of B-scan image, which is the horizontal position of target. We compute the vertical coordinates,which is the depth coordinates of target. Use the 1D Hough transform algorithm to estimate the equivalent wave velocity of the medium and calculate the actual depth of target. The results of simulation and experiment show that compared to energy statistics method the proposed method can position target more accurately.To indentify whether the target that have been positioned is void or not, we propose a new feature extraction method-- discrete wavelet transform- fractional Fourier transform(DWT-FRFT). The feature combined with support vector machine(SVM) classifiers to classify the signals of different material automatically. The results of simulation and experiment show that the proposed feature-based SVM system has good performances in classification accuracy compared to statistical and frequency domain feature-based SVM system.