Ground Penetrating Radar Data Inversion Using Dual-Input Convolutional Autoencoder for Ferroconcrete Inspection

Ground penetrating radar (GPR) is a non-destructive tool for exploring an object buried underground. Currently, GPR is also considered for reinforced concrete inspection. However, the image produced by GPR can not be easily interpreted. Besides, the large observation of building concrete inspection also motivates the researchers to fastening and easing radar image interpretation. Thus, this research proposes a new method to translate GPR scattering data image to its internal structure visualization. The proposed employs a convolutional autoencoder model using amplitude and phase radar data as input of the algorithm. As evaluation, in this stage, we perform numerical analysis by using finite-difference time-domain-based synthetic data that considers three cases: concrete with rebar, concrete with crack, and concrete with rebar and crack. All of those cases are simulated with randomized dimensions and positions that is possible in the real applications. Compared with the baseline method, our method shows superiority, especially in the semantic segmentation perspective. The parameter size of the proposed model is also much smaller, around one-third of the previous method. Therefore, the method is feasible enough to be implemented in real applications addressing an automatic internal structure reinforced concrete visulaization