Accelerating Constrained SIRTwith Applications in Tomographic Particle Image Reconstruction

Stefania Petra, Constantin Popa, Christoph SchnörrheiDOK 2009 · DOI

Abstract

This paper investigates a constrained version of the Simultaneous Iterative Reconstruction Technique (SIRT) through the lens of projected gradient methods. By establishing this connection, we assess the computational merits of this algorithm class and introduce acceleration procedures borrowed from numerical optimization. Specifically, we propose a spectral gradient steplength selection combined with a nonmonotone line search strategy to overcome the notoriously slow convergence of standard SIRT. We evaluate these schemes on algebraic image reconstruction models with sparsity constraints, focusing on tomographic particle image reconstruction. Our results demonstrate that the nonmonotone spectral projected gradient approach significantly outperforms traditional constrained SIRT across various constraining strategies.

Introduction

Tomographic Particle Image Velocimetry (TomoPIV) is a vital technique in fluid dynamics research, used to reconstruct 3D particle volumes from a limited set of 2D camera projections. However, this reconstruction is a "severely ill-posed" problem. Because the data is highly undersampled—with far more voxels than available measurement pixels—standard reconstruction techniques often struggle.

While the Simultaneous Iterative Reconstruction Technique (SIRT) is widely used due to its simplicity and parallelizable nature, it suffers from two primary limitations: extremely slow convergence (often requiring hundreds of thousands of iterations) and a tendency to produce dense solutions that do not reflect the sparse, discrete nature of physical particles.

Bridging SIRT and Numerical Optimization

The core contribution of this work is the formalization of SIRT as a projected gradient method. By framing SIRT as an optimization problem on a weighted least-squares objective, we can apply advanced optimization theories to improve its performance.

Constrained SIRT Framework

We generalize the SIRT iteration to include a constraining function $C$. This allows the reconstruction to be "steered" toward physically meaningful sets. Key constraints explored include:

  • The Positive Orthant: Ensuring all reconstructed intensities are non-negative.
  • The $\ell_1$-ball: Promoting sparsity in the solution.
  • The Simplex: A combination of positivity and $\ell_1$ constraints that proved particularly effective for particle reconstruction.

Spectral Acceleration (SPG)

To address the convergence bottleneck, we implement the Spectral Projected Gradient (SPG) method. This replaces the conservative constant steplength of traditional SIRT with the Barzilai-Borwein (BB) spectral steplength. The BB method uses information from previous iterations to mimic second-order (Hessian) behavior without the prohibitive computational cost of calculating the Hessian itself.

To ensure stability while maintaining speed, we employ a nonmonotone line search. This strategy allows the objective function to occasionally increase during iterations, preventing the algorithm from becoming trapped in local plateaus and accelerating the path to the global optimum.

Experimental Methodology

We evaluated the proposed methods using a simulated 2D TomoPIV model designed to replicate the challenges of 3D fluid imaging. The setup included:

  • Grid: A 2D volume consisting of 4,356 grid points.
  • Particles: Test cases featuring 5, 10, and 20 particles placed randomly or on-grid.
  • Sensors: Four 50-pixel cameras positioned at angles of $45^\circ, 15^\circ, -15^\circ,$ and $-45^\circ$.
  • Noise Levels: Gaussian-distributed noise ($\epsilon = 0, 0.05, 0.1$) was added to test the robustness of the reconstruction.

Key Results and Insights

Massive Computational Speedup

The most striking result was the disparity in convergence rates. In noiseless tests, the SPG algorithm reached the target accuracy in approximately 5,400 iterations. In contrast, the standard constrained SIRT required over 464,000 iterations to reach the same threshold. This represents a speedup of nearly two orders of magnitude.

The Efficacy of the Simplex

While simple positivity constraints were found to implicitly promote sparsity, the explicit use of simplex projections provided the best results. The simplex constraint consistently yielded the lowest iteration counts across all tested particle densities and provided superior denoising capabilities.

Robustness to Noise

Under significant noise ($\epsilon = 0.1$), the SPG method with simplex constraints maintained high reconstruction quality. Traditional SIRT methods either failed to converge within a practical timeframe or produced results with significant artifacts, highlighting the importance of the spectral approach in experimental settings where noise is inevitable.

Conclusion and Future Work

By treating SIRT as an optimization problem, we have unlocked significant performance gains for tomographic reconstruction. The combination of spectral steplengths and simplex constraints allows for rapid, sparse reconstruction of particle volumes.

Future research will focus on identifying the "support" (the set of non-zero voxels) earlier in the iteration process to enable active set methods. Additionally, while this study utilized 2D models, the authors anticipate that the transition to 3D volumes—though computationally demanding in terms of projection calculations—will follow the same mathematical principles established here.

Summary updated

Abstract

We investigate a constrained version of simultaneous iterative reconstruction techniques(SIRT) from the general viewpoint of projected gradient methods. This connection enable us to assess the computational merit of this algorithm class. We borrow a leaf from numerical optimization to cope with the slow convergence of projected gradient methods and propose an acceleration procedure based on the spectral gradient choice of steplength as in [2] and a nonmonotone strategy [17,4]. We compare these schemes and present numerical experiments on some algebraic image reconstruction models with sparsity constraints, with particular attention to tomographic particle image reconstruction. The performance of both constrained SIRT and nonmonotone spectral projected gradient approach is illustrated for several constraining strategies.
BibTeX
@inproceedings{petra2009accelerating,
  author = {Stefania Petra and Constantin Popa and Christoph Schnörr},
  title = {{Accelerating Constrained SIRTwith Applications in Tomographic Particle Image Reconstruction}},
  booktitle = {heiDOK},
  year = {2009},
  doi = {10.11588/heidok.00009477},
}
Figure