Unsupervised Deep Unrolling Networks for Phase Unwrapping

Zhile Chen, Yuhui Quan, Hui Ji

Presented at CVPR 2024, this research introduces U3Net, a novel framework designed to solve the phase unwrapping problem without the need for expensive ground-truth data. By combining the physics of phase wrapping with the efficiency of deep learning, U3Net establishes a new state-of-the-art for unsupervised phase unwrapping.

The Challenge of Phase Unwrapping

Phase Unwrapping (PU) is a fundamental step in various imaging technologies, including MRI, Interferometric Synthetic Aperture Radar (InSAR), and 3D depth sensing. Sensors typically measure phase values "wrapped" within the range $[-\pi, \pi)$. To recover the actual physical measurement, these values must be "unwrapped" by adding appropriate integer multiples of $2\pi$.

Traditional PU methods often rely on handcrafted heuristics that struggle with high noise. While supervised deep learning (DL) models have shown promise, they require massive amounts of ground-truth (GT) data, which is frequently unavailable in real-world scientific applications. Existing unsupervised methods, such as PUDIP, are "internal," meaning they require a time-consuming optimization process for every new image at test time, making them impractical for real-time deployment.

Comparison of learning paradigms: Supervised Learning (requires hard-to-collect GT), Internal Learning (time-consuming), and the proposed End-to-End Unsupervised Learning (bypassing both issues).

Methodology: The U3Net Framework

The authors propose U3Net (Unsupervised, Unrolling, and Unwrapping), which introduces three key innovations to bypass the need for ground truth while maintaining high performance:

  1. Gradient-Domain Learning: The core insight is that the gradients of wrapped phases are identical to those of the true phase, except at sparse "outlier" points where phase jumps occur. This allows the network to use wrapped gradients as noisy labels for training.
  2. Deep Unrolling Network (DUN): Rather than using a standard "black box" CNN, U3Net unrolls a variational optimization model into a neural network. This architecture explicitly includes a sub-network to estimate an error matrix ($E$), which identifies and neutralizes outliers that violate physical continuity conditions.
  3. Novel Loss Functions:
    • Noise-resistant self-reconstruction loss ($L_{sr}$): Employs a "recorruption" strategy to ensure the model learns the underlying signal rather than memorizing measurement noise.
  4. Self-distillation loss ($L_{sd}$): Uses intermediate predictions to provide pseudo-supervision, reducing ambiguity during training.
Extracted figure from page 4 of Unsupervised Deep Unrolling Networks for Phase Unwrapping

Key Results and Performance

U3Net was evaluated against synthetic datasets (MoGR and RME) and real-world-inspired InSAR phase data. The results demonstrate that U3Net is both highly accurate and exceptionally fast:

  • Competitive Accuracy: In synthetic tests, U3Net achieved the best Normalized Root Mean Squared Error (NRMSE) in 8 out of 10 noise settings. On InSAR data, it outperformed several supervised models, including PUNet and PU-GAN.
  • Unprecedented Speed: As an "external" learning model, U3Net learns from a general dataset once. After training, it can unwrap a $256 \times 256$ image in approximately 10 milliseconds—roughly 10,000 times faster than the internal unsupervised method PUDIP.
  • Efficiency: The model is lightweight, with only 0.74M parameters, making it suitable for resource-constrained environments.

Conclusion and Future Directions

U3Net represents a significant leap forward by providing an end-to-end unsupervised solution that rivals supervised performance. While current iterations assume some knowledge of noise statistics for the "recorruption" loss, future work will focus on making the model robust to unknown or non-Gaussian noise profiles common in complex real-world environments. By bypassing the ground-truth requirement, U3Net opens new possibilities for high-speed, high-accuracy phase unwrapping in scientific and industrial imaging.

Summary updated

Abstract

Phase unwrapping (PU) is a technique to reconstruct original phase images from their noisy wrapped counterparts, finding many applications in scientific imaging. Although supervised learning has shown promise in PU, its utility is limited in ground-truth (GT) scarce scenarios. This paper presents an unsupervised learning approach that eliminates the need for GTs during end-to-end training. Our approach leverages the insight that both the gradients and wrapped gradients of wrapped phases serve as noisy labels for GT phase gradients, along with sparse outliers induced by the wrapping operation. A recorruption-based self-reconstruction loss in the gradient domain is proposed to mitigate the adverse effects of label noise, complemented with a self-distillation loss for improved generalization. Additionally, by unfolding a variational model of PU that utilizes wrapped gradients of wrapped phases for its data-fitting term, we develop a deep unrolling network that encodes physics of phase wrapping and incorporates special treatments on outliers. In the experiments on three types of phase data, our approach outperforms existing GT-free methods and competes well against the supervised ones.
BibTeX
@inproceedings{chen2024unsupervised,
  author = {Zhile Chen and Yuhui Quan and Hui Ji},
  title = {{Unsupervised Deep Unrolling Networks for Phase Unwrapping}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2024},
  doi = {10.1109/cvpr52733.2024.02379},
}
Figure