Speaker
Description
Image reconstruction in muon scattering tomography is a complex task, and finding suitable reconstruction algorithms for a given application often requires compromises: Simple approaches, such as the Point of Closest Approach (PoCA) or Angle Statistics Reconstruction (ASR) algorithm, require little computational effort but are inherently limited in achievable image quality due to their underlying simplifications. In contrast, statistical approaches can provide significantly improved results; however, their implementation is numerically complex and requires substantial computing resources.
In this work, we present a statistical maximum-likelihood reconstruction framework that enables both accurate and computationally efficient muon image reconstruction. The method is based on a per-voxel likelihood formulation of the muon scattering information, quantifying the agreement between a candidate density map and the detector data. The likelihood is maximized directly using gradient descent optimization, treating the voxel densities as free parameters and therefore yielding an optimal density map.
The gradient descent optimization is performed via automatic differentiation, a method that allows for efficient gradient calculations that are free of uncertainties. To ensure scalability to large datasets and compatibility with limited computational resources, the implementation leverages parallelization, GPU acceleration, and batched processing.
This contribution introduces the theoretical foundations and practical implementation of the proposed gradient descent reconstruction algorithm. Based on example scenes with varying complexities, the performance of the algorithm is compared to traditional algorithms with respect to reconstruction quality, runtime, and memory consumption.