Speaker
Description
Across industry, muography is becoming an increasingly prevalent next-generation non-destructive evaluation technique. A key limitation of muon tomography, however, remains the low natural muon flux, which leads to long acquisition times, noisy reconstructions and artefacts that complicate interpretation.
This work explores the application of deep learning to post-reconstruction muon tomography data to enhance image quality and feature recovery. Using simulated datasets, we demonstrate that U-Nets and conditional generative adversarial networks can be deployed for predictive upsampling and detailed semantic segmentation. Structural Similarity Index Measure (SSIM) and Peak Signal-to-Noise Ratio (PSNR) evaluations indicate that machine learning can reduce effective imaging times for reinforced concrete structures from weeks to days. Segmentation performance shows similar time-dependent improvements, while also mitigating z-plane smearing effects and improving overall interpretability.
Comparable learning-based approaches have been applied in other imaging domains to achieve substantial reductions in the number of events required to produce target results, suggesting broader applicability of these techniques. Overall, this work demonstrates how machine learning can both accelerate data acquisition requirements and enhance reconstruction quality, supporting the wider industrial adoption of muographic imaging.