Thursday 20 March 2025
Deep learning has been making waves in various fields, and its application in muography, a non-invasive imaging technique used to detect internal structures of concrete structures, is no exception. Muography uses cosmic-ray muons to create three-dimensional density maps of scanned volumes. However, the technology’s reliance on naturally occurring muons results in prolonged acquisition times, noisy reconstructions, and challenges in image interpretation.
Researchers have been working to address these limitations by developing deep learning models that can enhance muography images and improve their interpretability. One such approach is the use of conditional Wasserstein generative adversarial networks with gradient penalty (cWGAN-GP). This model is designed to perform predictive upsampling of undersampled muography images, which can significantly reduce acquisition times.
The cWGAN-GP model was trained on a comprehensive dataset generated using Geant4 Monte-Carlo simulations, which reflect realistic civil infrastructure scenarios. The results are impressive: the 1-day sampled images were able to match the perceptual qualities of 21 baseline images in terms of structural similarity index measure (SSIM) and peak signal-to-noise ratio (PSNR). This represents a significant reduction in required sampling time for achieving high-quality muography images.
But that’s not all. The researchers also developed a second cWGAN-GP model, trained for semantic segmentation, to assess the upsampling model’s impact on each feature within the concrete samples. This model was able to accurately identify rebar grids and tendon ducts embedded in the concrete, as well as mitigate z-plane smearing artifacts caused by the muography’s inherent inverse imaging problem.
The use of attention mechanisms in the cWGAN-GP models is particularly noteworthy. Unlike traditional convolutional neural networks (CNNs), which are limited to processing local features, attention mechanisms enable the model to capture long-range dependencies and spatial relationships between tokens. This allows the model to account for location dependence and improve its ability to identify objects within the concrete structures.
The potential applications of this technology are vast. Muography could be used to inspect and monitor built infrastructure, such as bridges and buildings, more efficiently and effectively than traditional methods. It could also be used to detect defects and anomalies in real-time, reducing the risk of catastrophic failures.
However, there is still much work to be done before muography can become a widely adopted technology. The researchers acknowledge that improving air void detection and developing more advanced segmentation capabilities are key areas of focus for future development.
Cite this article: “Enhancing Muography Images with Deep Learning for Concrete Structure Inspection”, The Science Archive, 2025.
Muography, Deep Learning, Image Enhancement, Concrete Structures, Non-Invasive Imaging, Cosmic-Ray Muons, Generative Adversarial Networks, Wasserstein Distance, Gradient Penalty, Semantic Segmentation
