Enhancing Industrial X-ray Tomography by Data-Centric Statistical Methods

Jarkko Suuronen, Muhammad F. Emzir, Sari Lasanen, Simo Särkkä, Lassi Roininen

Introduction

Industrial X-ray tomography is a critical tool for non-destructive testing and analysis across diverse sectors, including the sawmill industry, oil and gas exploration, and geotechnical engineering. Whether identifying metallic inclusions in timber or characterizing the pore structure of rock drill cores, the fidelity of the tomographic reconstruction is paramount. Traditional reconstruction techniques, such as Filtered Back-Projection (FBP) or Gaussian-based Bayesian methods, often struggle with industrial targets characterized by sharp boundaries and high-contrast features. These methods typically assume a degree of spatial smoothness that does not align with physical reality, leading to blurred edges and artifacts, particularly when data is sparse or noisy. This article explores the application of non-Gaussian Bayesian priors to address these limitations, promoting edge preservation and robust uncertainty quantification.

Methodology: Bayesian Inverse Problems and Statistical Priors

The reconstruction process is framed as a Bayesian statistical inverse problem, where the goal is to determine the internal structure of an object from 2D X-ray projections. The choice of a "prior" distribution is central to this approach, as it encodes assumptions about the object's properties. This study compares four distinct priors:

  1. Gaussian Priors: These favor overall smoothness and serve as the traditional baseline.
  2. Total Variation (TV) Priors: These are designed to promote sharp edges but lack mesh invariance, meaning results can vary with the resolution of the computational grid.
  3. Besov Priors: Wavelet-based priors that provide mesh invariance but may introduce "blocky" artifacts.
  4. Cauchy Priors: Heavy-tailed distributions that excel at capturing sharp transitions and high-contrast features, such as metallic inclusions.

To handle the high-dimensional nature of these problems—often involving between 100,000 and 1,000,000 unknowns—the researchers employed Maximum A Posteriori (MAP) estimation via the limited-memory BFGS (L-BFGS) optimization algorithm for rapid point estimates. For comprehensive uncertainty quantification (UQ), they utilized advanced Markov Chain Monte Carlo (MCMC) methods, specifically a no-U-turn variant of Hamiltonian Monte Carlo (HMC-NUTS) and adaptive Metropolis-within-Gibbs algorithms.

Experimental Setup and Datasets

The efficacy of these statistical methods was evaluated using two primary synthetic benchmarks simulating industrial challenges:

  • Sawmill Log Tomography: A high-contrast 3D model of a log containing knots, rot, and metallic pieces. This represents a "worst-case" scenario for traditional smoothing-based reconstruction.
  • Drill-core Rock Sample: A lower-contrast target focused on identifying complex pore structures and mineralized soil within a rock core.

Both datasets were tested under varying degrees of data sparsity, using 90, 30, and 10 measurement angles to assess performance under extreme conditions.

Results and Discussion

The experimental results demonstrate the clear superiority of non-Gaussian priors in industrial contexts. In the log tomography test with 90 angles, TV and Cauchy priors achieved Peak Signal-to-Noise Ratios (PSNR) of 36.6 dB and 36.0 dB, respectively, significantly outperforming the Gaussian prior's 33.7 dB.

The Cauchy prior proved exceptionally resilient to data sparsity. Even when reduced to only 10 measurement angles, it successfully identified metallic inclusions that were entirely lost in Gaussian reconstructions. Qualitatively, non-Gaussian methods suppressed the "streaking" artifacts typical of sparse-angle CT, producing clearer, more interpretable images.

Furthermore, the study confirms the viability of HMC-NUTS for high-dimensional uncertainty quantification. By providing "error bars" for individual pixels, these methods allow engineers to assess the reliability of a reconstruction, which is vital for critical decision-making in industrial processes.

Conclusion and Future Directions

While non-Gaussian Bayesian methods offer significant improvements in reconstruction quality and reliability, computational demands remain a challenge. MAP estimates are sufficiently fast for many applications, but full MCMC-based uncertainty quantification is currently too slow for real-time production environments like high-speed sawmills. Additionally, the heavy-tailed nature of Cauchy priors can complicate MCMC sampling.

Future research directions include the development of "Deep Bayesian Inversion" techniques. By integrating robust statistical priors with machine learning, it may be possible to achieve the computational speed required for industrial production while maintaining the rigorous physical grounding and uncertainty quantification provided by Bayesian frameworks.

Summary updated

Abstract

X-ray tomography has applications in various industrial fields such as sawmill industry, oil and gas industry, chemical engineering, and geotechnical engineering. In this article, we study Bayesian methods for the X-ray tomography reconstruction. In Bayesian methods, the inverse problem of tomographic reconstruction is solved with help of a statistical prior distribution which encodes the possible internal structures by assigning probabilities for smoothness and edge distribution of the object. We compare Gaussian random field priors, that favour smoothness, to non-Gaussian total variation, Besov, and Cauchy priors which promote sharp edges and high-contrast and low-contrast areas in the object. We also present computational schemes for solving the resulting high-dimensional Bayesian inverse problem with 100,000-1,000,000 unknowns. In particular, we study the applicability of a no-U-turn variant of Hamiltonian Monte Carlo methods and of a more classical adaptive Metropolis-within-Gibbs algorithm for this purpose. These methods also enable full uncertainty quantification of the reconstructions. For faster computations, we use maximum a posteriori estimates with limited-memory BFGS optimisation algorithm. As the first industrial application, we consider sawmill industry X-ray log tomography. The logs have knots, rotten parts, and even possibly metallic pieces, making them good examples for non-Gaussian priors. Secondly, we study drill-core rock sample tomography, an example from oil and gas industry. We show that Cauchy priors produce smaller number of artefacts than other choices, especially with sparse high-noise measurements, and choosing Hamiltonian Monte Carlo enables systematic uncertainty quantification.
BibTeX
@inproceedings{suuronen2020enhancing,
  author = {Jarkko Suuronen and Muhammad F. Emzir and Sari Lasanen and Simo Särkkä and Lassi Roininen},
  title = {{Enhancing Industrial X-ray Tomography by Data-Centric Statistical Methods}},
  booktitle = {University},
  year = {2020},
}
Figure