AUTHOR=Huang Zhiyong , Xu Guoyuan , Tang Jiaming , Yu Huayang , Wang Duanyi TITLE=Research on Void Signal Recognition Algorithm of 3D Ground-Penetrating Radar Based on the Digital Image JOURNAL=Frontiers in Materials VOLUME=9 YEAR=2022 URL=https://www.frontiersin.org/journals/materials/articles/10.3389/fmats.2022.850694 DOI=10.3389/fmats.2022.850694 ISSN=2296-8016 ABSTRACT=

The three-dimensional ground-penetrating radar system is an effective method to detect road void disease. Ground penetrating radar image interpretation has the characteristics of multi-solution, long interpretation period, and high professional requirements of processors. In recent years, researchers have put forward solutions for automatic interpretation of ground-penetrating radar images, including automatic detection algorithm for subgrade diseases based on support vector machines, etc., but there are still some shortcomings such as training models with a large amount of data or setting parameters. In this article, a three-dimensional ground-penetrating radar void signal recognition algorithm based on the digital image is proposed, and the algorithm uses digital images to characterize radar signals. With the help of digital image processing methods, the images are processed by binarization, corrosion, expansion, connected area inspection, fine length index inspection, and three-dimensional matching inspection, so as to identify and determine the void signals and extract the void area volume index. The algorithm has been verified by laboratory tests and engineering projects, and the results show that the void identification algorithm can accurately identify the void area position; the error level between the measured values and the measured values of length, width, buried depth, and area is between 2.2 and 17.3%, and the error is generally within the engineering acceptance range. The volume index calculated by the algorithm has a certain engineering application value; compared with the support vector machine, the regressive convolution neural network, and other recognition methods, it has the advantage of not needing a large amount of data to train or modify parameters.