Application of Deep Learning in Ground Penetrating Radar Image Recognition

Abstract

In the traditional ground-penetrating radar image recognition methods, the recognition accuracy is not high, the recognition of complex targets is difficult, the recognition process is cumbersome, and the structural characteristics reflected by the measured data cannot be accurately judged. In this paper, a Faster R-CNN target detection algorithm in deep learning is applied to GPR image recognition. Firstly, based on the three-phase discrete random medium model, GPRmax software is used for forwarding simulation, and GPR images of different earth-rock dam disease types are obtained by changing the position, shape, and range of different disease types, and GPR image data sets are expanded using horizontal flipping, random erasing, scale transformation and adding random noise. Under the framework of TensorFlow, based on the Faster R-CNN target detection algorithm, the training and testing of the earth-rock dam disease target detection model are completed. The recognition performance on the test set is analyzed from three indexes of recall rate, precision rate and average precision rate, and evaluation indexes such as accuracy-recall rate curve. The results show that the average accuracy of the constructed target detection model for earth-rock dam disease recognition is over 90%. Among them, the recognition effect of concentrated leakage, scattered leaching, and cracks is the best. Finally, the effectiveness of the target detection model established in this paper is verified by the measured data of an abandoned slag slope.