Abstract
This paper presents a new algorithm, Accelerated Wirtinger Flow (AWF), for ptychographic image reconstruction from phaseless diffraction pattern measurements. AWF is based on combining Nesterov's acceleration approach with Wirtinger gradient descent. Theoretical results enable prespecification of all AWF algorithm parameters, with no need for computationally-expensive line searches and no need for manual parameter tuning. AWF is evaluated in the context of simulated X-ray ptychography, where we demonstrate fast convergence and low per-iteration computational complexity. We also show examples where AWF reaches higher image quality with less computation than classical algorithms. AWF is also shown to have robustness to noise and probe misalignment.
Introduction
In nano-scale imaging, such as the study of protein complexes or 10nm transistors, traditional lenses are often impractical due to the difficulty of manufacturing them for short wavelengths like X-rays. To overcome this, researchers employ ptychography. In this technique, a coherent beam, or "probe," illuminates a sample at various positions on a grid, and the resulting diffraction patterns are recorded.
A significant challenge in ptychography is that detectors only record the intensity of the light, losing the phase information. Reconstructing the image thus becomes a "phase retrieval" problem—a large-scale, non-convex optimization task. Traditional algorithms often suffer from slow convergence, high computational costs due to per-iteration "line searches," and the need for extensive manual parameter tuning.
Methodology: Accelerated Wirtinger Flow (AWF)
The proposed Accelerated Wirtinger Flow (AWF) algorithm addresses these challenges by combining two key optimization concepts:
- Wirtinger Gradient Descent: This framework allows for the calculation of gradients for real-valued functions involving complex variables, which is essential since both images and light waves are complex-valued.
- Nesterov’s Acceleration: This "momentum" technique utilizes information from previous iterations to accelerate the update process toward the optimal solution.
A core contribution of this work is the derivation of a fixed step-size ($\bar{\mu}$). Typically, ptychographic reconstruction requires a computationally expensive line search in each iteration to determine the optimal step size, involving multiple Fast Fourier Transforms (FFTs). The authors prove that a step size equal to the reciprocal of the largest eigenvalue of the system matrix ($1/\lambda_{max}(\mathbf{A}^H\mathbf{A})$) ensures convergence.
As shown in the figure above, this value is directly related to the probe's illumination pattern and is remarkably simple to calculate, significantly reducing the computational overhead per iteration.
Experimental Evaluation
The AWF algorithm was tested using high-fidelity simulations of synchrotron X-ray experiments. The datasets included:
- Integrated Circuit (IC) Image: A 500x500 pixel complex-valued image simulating a 3D microchip projection.
- Gold Beads: A standard test image featuring gold spheres on a membrane.
- Noise and Error Models: Simulations included Poisson noise to mimic photon counting and random probe misalignment to account for mechanical stage errors.
Key Results and Discussion
AWF demonstrated several advantages over existing methods:
- Computational Efficiency: By eliminating line searches, AWF reduces the number of FFTs required per iteration. When measured against the total number of FFT computations, AWF outperformed standard Wirtinger Flow (WF), Conjugate Gradient (CGM), ePIE, and Difference Map (DM).
- Rapid Convergence: AWF reached target error levels in significantly fewer iterations than standard WF.
- Robustness: The algorithm maintained high image quality under significant Poisson noise ($10^7$ photons per probe position). Furthermore, it showed greater resilience to probe misalignment compared to the Difference Map algorithm, which tended to become unstable.
- Ease of Use: The theoretical pre-specification of parameters allows AWF to function effectively "out of the box," removing the need for expert manual tuning.
The choice of a cost function based on the amplitude of the diffraction pattern, rather than intensity, was also found to enhance convergence speed and noise robustness.
Conclusion and Future Work
Accelerated Wirtinger Flow provides a robust and efficient solution for ptychographic reconstruction. While the current study assumes a known probe shape, preliminary results suggest AWF can be integrated with probe-update strategies. Future work will focus on developing a rigorous theoretical framework for simultaneous probe and image reconstruction using acceleration techniques and validating the algorithm with experimental data from real-world synchrotron facilities.
Summary updated