Fast and Robust Phase Retrieval via Deep Expectation-Consistent Approximation

Saurav K. Shastri,Philip Schniter
2024-07-13
Abstract:Accurately recovering images from phaseless measurements is a challenging and long-standing problem. In this work, we present "deepECpr," which combines expectation-consistent (EC) approximation with deep denoising networks to surpass state-of-the-art phase-retrieval methods in both speed and accuracy. In addition to applying EC in a non-traditional manner, deepECpr includes a novel stochastic damping scheme that is inspired by recent diffusion methods. Like existing phase-retrieval methods based on plug-and-play priors, regularization by denoising, or diffusion, deepECpr iterates a denoising stage with a measurement-exploitation stage. But unlike existing methods, deepECpr requires far fewer denoiser calls. We compare deepECpr to the state-of-the-art prDeep (Metzler et al., 2018), Deep-ITA (Wang et al., 2020), and Diffusion Posterior Sampling (Chung et al., 2023) methods for noisy phase-retrieval of color, natural, and unnatural grayscale images on oversampled-Fourier and coded-diffraction-pattern measurements and find improvements in both PSNR and SSIM with 5x fewer denoiser calls.
Computer Vision and Pattern Recognition,Information Theory
What problem does this paper attempt to address?
### What problem does this paper attempt to solve? This paper aims to solve the difficult problem of accurately recovering images from phaseless measurements. Specifically, the author proposes a new method named **deepECpr**, which combines the Expectation - Consistent (EC) approximation with a deep denoising network to surpass the speed and accuracy of existing phase retrieval (PR) methods. #### Specific problem description 1. **Challenges of phaseless measurements**: - In many application scenarios (such as optics, holography, X - ray crystallography, diffraction imaging, non - line - of - sight imaging, astronomy, computational biology, speech recognition, etc.), due to technical limitations, it is impossible to measure phase information and only amplitude information can be obtained. - This makes it very difficult to recover the original signal or image from these phaseless measurements. Even when the linear transformation \( A \) is full - rank and there is no noise, there are still multiple possible solutions \( x \) such that \( y = |Ax|+w \) holds. 2. **Limitations of existing methods**: - **Classical methods**: Such as the Gerchberg - Saxton (GS) and Hybrid Input - Output (HIO) algorithms. Although they are simple and fast, their output quality is not as good as that of modern methods. - **Gradient - based methods**: Solve non - convex optimization problems by minimizing the negative log - likelihood (NLL). However, due to non - convexity, they are prone to getting trapped in local minima. - **Convex relaxation methods**: Such as PhaseLift and PhaseMax. Although they are theoretically effective, their computational complexity in high - dimensional spaces is high and they are not suitable for practical imaging applications. - **Prior - knowledge - based methods**: Such as Plug - and - Play (PnP) and Regularization by Denoising (RED). Although they can use deep - learning denoisers to improve performance, they still need to call the denoiser multiple times and are less efficient. #### Innovations of deepECpr 1. **Application of the Expectation - Consistent (EC) framework**: - deepECpr uses the EC approximation framework to estimate \( z \) instead of directly estimating \( x \), thus avoiding the dependence on the large random matrix \( A \). 2. **Novel random damping scheme**: - A random damping mechanism inspired by diffusion methods is introduced, which further alleviates the model assumption deviation and improves convergence and stability. 3. **Efficient iterative process**: - deepECpr achieves a faster convergence speed and higher accuracy with fewer denoiser calls. Compared with existing methods (such as prDeep, Deep - ITA, and DPS), it shows significant advantages in PSNR and SSIM metrics. 4. **Extensive experimental verification**: - The paper verifies the superior performance of deepECpr in processing color images, natural grayscale images, and non - natural grayscale images through a large number of numerical experiments, especially under oversampled Fourier measurements (OSF) and coded diffraction pattern measurements (CDP). In conclusion, this paper proposes a new phase - recovery method, aiming to solve the deficiencies of existing methods in terms of speed, accuracy, and robustness, and provides a more effective solution for image recovery from phaseless measurements.