Three-Dimensional Microwave Imaging: Fast and Accurate Computations with Block Resolution Algorithms

Corentin Friedrich, Sébastien Bourguignon, Jérôme Idier, Yves GoussardSensors 2020 · OpenReview

TL;DR

This paper introduces a novel "block-BiCGStab" algorithm to significantly speed up 3D microwave imaging reconstructions. By solving multiple linear systems simultaneously, it achieves up to a 2x speedup compared to traditional methods, especially for challenging scenarios with high frequencies or strong object contrasts.

What Problem Does This Paper Address?

Microwave imaging aims to map the internal dielectric properties of an object by analyzing how it scatters electromagnetic waves. This is crucial for applications like medical imaging and non-destructive testing. However, reconstructing a 3D image from scattered field data is computationally intensive. The core challenge lies in solving numerous large, coupled linear systems that arise from the forward scattering model, especially when dealing with high frequencies (which provide more detail but are harder to model) or objects with stark differences in material properties (high contrast). Existing methods often struggle with the sheer computational cost, limiting their practical application.

Methods

The paper proposes a "block-BiCGStab" algorithm, an adaptation of the well-established BiCGStab iterative solver. The key innovation is its ability to solve multiple, related linear systems concurrently, rather than one by one. In microwave imaging, each "system" corresponds to a different incident electromagnetic field (i.e., a different source). Since these systems share a common underlying matrix, solving them in a "block" fashion leverages this shared structure for efficiency.

The authors also highlight the importance of proper initialization for the block-BiCGStab algorithm. They found that initializing with a slightly noisy version of the incident field, rather than just the incident field itself, significantly improves convergence and prevents divergence, especially in challenging scenarios. This is a crucial practical detail for the algorithm's success.

Datasets

The paper uses two main sources of data:

  • Synthetic Data: A simulated 3D object composed of two nested cubes was created. This dataset allowed for controlled experiments to evaluate the algorithm's performance under varying frequencies and contrast levels. The object was discretized into 30x30x30 voxels.
  • Fresnel Database Objects: The authors tested their method on five objects from the publicly available Fresnel database. These are real-world measurement datasets representing challenging scenarios, including objects like "TwoSpheres" and "Cylinder," which are known to be difficult to reconstruct.

Key Insights and Ideas

  • The Power of Parallelism (Implicit): While not a direct parallel computing implementation, the block-BiCGStab algorithm effectively exploits the inherent structure of the problem to solve multiple systems simultaneously, mimicking a form of parallelism. This is a crucial insight for tackling large-scale inverse problems.
  • Initialization is Key: The paper demonstrates that a seemingly small detail like how you initialize the iterative solver can have a dramatic impact on convergence and accuracy, especially for complex problems. The noisy initialization strategy is a practical workaround for a common pitfall.
  • Difficulty Drives Benefit: The most significant speedups and performance improvements are observed in the most challenging scenarios (high frequencies, high contrast). This suggests the block approach is not just a minor optimization but a fundamental improvement for tackling the hardest problems in microwave imaging.

Key Results

  • Significant Speedup: The block-BiCGStab algorithm achieved up to a 2x reduction in computation time compared to the standard sequential BiCGStab. For the most difficult object in the Fresnel database (Cylinder), the reconstruction time dropped from 60 hours to 28.3 hours.
  • Improved Convergence: For difficult problems, the block-BiCGStab algorithm converged in 40% fewer iterations than the sequential version when properly initialized.
  • Accurate Reconstructions: The method successfully reconstructed both synthetic and real-world objects, including challenging ones from the Fresnel database. The shapes and contrasts of objects were well-retrieved, with results comparable to or better than previous state-of-the-art methods.

Limitations and Future Work

  • Computational Cost: While significantly improved, 3D microwave imaging reconstruction remains computationally demanding. The paper notes that the block-BiCGStab algorithm still requires substantial computation.
  • Parallelization: The authors suggest that further improvements could be achieved through explicit parallel implementation of the block-BiCGStab algorithm, potentially leveraging multicore architectures, GPUs, or High-Performance Computing (HPC) clusters.
  • Generalizability: The paper focuses on additive regularization. While the authors believe their block strategy could be adapted to multiplicative regularization, further study would be needed. They also suggest exploring the algorithm's performance in different geometrical configurations and with varying discretization schemes.
  • Tuning: Specific tuning of the block-BiCGStab algorithm might be required for different problem complexities, and the strategies presented in the paper could guide this tuning process.

Summary updated

Abstract

This paper considers the microwave imaging reconstruction problem, based on additive penalization and gradient-based optimization. Each evaluation of the cost function and of its gradient requires the resolution of as many high-dimensional linear systems as the number of incident fields, which represents a large amount of computations. Since all such systems involve the same matrix, we propose a block inversion strategy, based on the block-biconjugate gradient stabilized (BiCGStab) algorithm, with efficient implementations specific to the microwave imaging context. Numerical experiments performed on synthetic data and on real measurements show that savings in computing time can reach a factor of two compared to the standard, sequential, BiCGStab implementation. Improvements brought by the block approach are even more important for the most difficult reconstruction problems, that is, with high-frequency illuminations and/or highly contrasted objects. The proposed reconstruction strategy is shown to achieve satisfactory estimates for objects of the Fresnel database, even on the most contrasted ones.
BibTeX
@inproceedings{friedrich2020three,
  author = {Corentin Friedrich and Sébastien Bourguignon and Jérôme Idier and Yves Goussard},
  title = {{Three-Dimensional Microwave Imaging: Fast and Accurate Computations with Block Resolution Algorithms}},
  booktitle = {Sensors},
  year = {2020},
  url = {https://openreview.net/forum?id=nsvnQtPYVN},
}
Figure