Recently, a research group led by Prof. LI Baopeng and Prof. MA Caiwen from the Xi’an Institute of Optics and Precision Mechanics (XIOPM), Chinese Academy of Sciences, developed a groundbreaking approach to address the challenges of noise robustness and computational complexity in traditional Fourier ptychography algorithms. This advancement holds significant promise for high-resolution imaging in dynamic real-world environments.
The research was published in Computational Visual Media.
Fourier ptychography, a computational imaging technique that synthesizes high-resolution images from multiple low-resolution inputs, faces limitations in handling system noise and phase retrieval. While deep learning methods have improved reconstruction speed, their "black-box" nature limits interpretability and performance in complex scenarios.
To overcome these issues, the team designed MDFP-Net, a model-driven neural network that unfolds a custom optimization algorithm into interpretable neural modules. The network alternates between real and complex domains, enabling simultaneous recovery of intensity and phase information – a critical capability for applications like cellular imaging and material analysis.
In experiments, two standalone components were central to the design: (1) Z-Net – Updates complex spectral data using gradient descent aligned with FP physics; (2) X-Net – Refines the reconstructed image via a learned proximal operator, enhancing noise resistance. The system demonstrated remarkable stability, achieving 31.45 dB peak signal-to-noise ratio (PSNR) on the prDeep dataset under noise-free conditions, surpassing traditional methods by 1.36 dB. Even under severe noise, MDFP-Net maintained 28.13 dB PSNR, outperforming conventional algorithms by 2.4 dB.
Additionally, the team validated the network using a custom-built, long-distance reflection Fourier ptychography system. Reconstructed images of resolution charts and coins showed 60% fewer artifacts compared to existing methods, highlighting its practicality for macroscopic imaging.
“MDFP-Net bridges the gap between physics-based models and data-driven learning, offering both performance and interpretability, and it could redefine computational microscopy in fields requiring high precision, such as medical diagnostics and aerospace inspection," said LI Baopeng.
(published April 25, 2025)
Fig. The proposed MDFP-Net. (Image by XIOPM)
Download：