Published: 2023

Publication Name: Denoising of GPR B-scan images using discrete wavelet transform

Publication URL: https://doi.org/10.1117/12.2657741

Abstract:

TIDC Project 3.19

 

Subsurface inspection of concrete structures using electromagnetic (EM) sensors such as ground penetrating radar (GPR) and synthetic aperture radar (SAR) is a field applicable approach for critical civil infrastructure systems. Compared to other nondestructive inspection/evaluation/testing techniques, EM waves can penetrate the surface of concrete structures and travel inside the subsurface of concrete structures to generate backscattering signals from a subsurface target (e.g., corroded steel rebar, delamination/cracking) for condition assessment. However, variations in the EM property of concrete and unpredictable EM background noises can contribute to the difficulties of image interpretation. Denoising of radar images is a necessary step before engineers can perform quantitative assessment of the images. The objective of this paper is to denoise GPR images using discrete wavelet transform (DWT). Four concrete panel specimens (30-by-30-by-3.5 cm3 ) were prepared with three artificial cracks (CNC, CNCD, and CNCW) of known dimensions and subjected to B-scan inspection using a 1.6 GHz GPR sensor (StructureScan Mini, GSSI). Level five Daubechies wavelet was used in processing all GPR B-scan images for its capability of detecting high frequency components in this study. The purpose of image denoising was to reveal a clear hyperbolic pattern by eliminating undesired local maximum and minimum points. After each image processing, horizontal, vertical, and diagonal details were generated. Four different denoising schemes were considered; i) without all details, ii) horizontal detail only, iii) vertical detail only, and iv) diagonal detail only. Denoising was repeated in five steps in each scheme. Performance of denoising was evaluated by the number of local maximum and minimum points and their geometric pattern. From our results, it was found that, for some type of artificial crack (CNCD), the hyperbolic pattern can be clearly revealed after one step of denoising, regardless of the denoising scheme. Among four denoising schemes, the best scheme is the one with the approximation coefficients and the diagonal detail coefficients. It was also found that including horizontal detail can introduce high frequency artifacts, resulting in an over-denoised image.