Muographic Image Upsampling with Machine Learning for Built Infrastructure Applications
William O'Donnell, David Mahon, Guangliang Yang, Simon Gardner
TL;DR
The paper addresses the challenge of slow muography-based inspection of reinforced concrete by proposing a two-model deep learning approach trained on Geant4-simulated data. A cWGAN-GP upsamples undersampled muography images, while a second model performs semantic segmentation to quantify how upsampling affects feature visibility such as rebar grids and tendon ducts, and to mitigate z-plane shadowing. The results show that 1-day undersampled images can achieve perceptual quality comparable to 21 days and noise levels akin to 31 days, with segmentation revealing strong gains for tendon ducts and moderate gains for rebar, though air voids remain difficult due to class imbalance. The findings demonstrate meaningful gains in acquisition speed and image interpretability, indicating a viable path toward practical muography for infrastructure monitoring, with future work focusing on 3D context, defect detection, and real-world validation.
Abstract
The civil engineering industry faces a critical need for innovative non-destructive evaluation methods, particularly for ageing critical infrastructure, such as bridges, where current techniques fall short. Muography, a non-invasive imaging technique, constructs three-dimensional density maps by detecting interactions of naturally occurring cosmic-ray muons within the scanned volume. Cosmic-ray muons provide deep penetration and inherent safety due to their high momenta and natural source. However, the technology's reliance on this source results in constrained muon flux, leading to prolonged acquisition times, noisy reconstructions and image interpretation challenges. To address these limitations, we developed a two-model deep learning approach. First, we employed a conditional Wasserstein generative adversarial network with gradient penalty (cWGAN-GP) to perform predictive upsampling of undersampled muography images. Using the Structural Similarity Index Measure (SSIM), 1-day sampled images matched the perceptual qualities of a 21-day image, while the Peak Signal-to-Noise Ratio (PSNR) indicated noise improvement equivalent to 31 days of sampling. A second cWGAN-GP model, trained for semantic segmentation, quantitatively assessed the upsampling model's impact on concrete sample features. This model achieved segmentation of rebar grids and tendon ducts, with Dice-Sørensen accuracy coefficients of 0.8174 and 0.8663. Notably, it could mitigate or remove z-plane smearing artifacts caused by muography's inverse imaging problem. Both models were trained on a comprehensive Geant4 Monte-Carlo simulation dataset reflecting realistic civil infrastructure scenarios. Our results demonstrate significant improvements in acquisition speed and image quality, marking a substantial step toward making muography more practical for reinforced concrete infrastructure monitoring applications.