DLMMPR:Deep Learning-based Measurement Matrix for Phase Retrieval

TL;DR

This work addresses the efficiency and robustness of phase retrieval by learning a measurement matrix within an end-to-end deep network. DLMMPR unrolls ISTA-like updates into a trainable architecture that jointly optimizes the measurement matrix and reconstruction steps, embedding a subgradient descent module and a proximal mapping module with a learnable nonlinear transform to enforce sparsity. Across extensive experiments on CDP-based measurements and multiple noise regimes, DLMMPR consistently outperforms state-of-the-art methods in PSNR and SSIM, demonstrating improved reconstruction quality and resilience to noise. The approach has practical implications for high-fidelity, efficient imaging in applications like diffraction, microscopy, and remote sensing, with future work focusing on scalability and real-time deployment.

Abstract

This paper pioneers the integration of learning optimization into measurement matrix design for phase retrieval. We introduce the Deep Learning-based Measurement Matrix for Phase Retrieval (DLMMPR) algorithm, which parameterizes the measurement matrix within an end-to-end deep learning architecture. Synergistically augmented with subgradient descent and proximal mapping modules for robust recovery, DLMMPR's efficacy is decisively confirmed through comprehensive empirical validation across diverse noise regimes. Benchmarked against DeepMMSE and PrComplex, our method yields substantial gains in PSNR and SSIM, underscoring its superiority.

Figures (7)

• Figure 1: Coded Diffraction Patterns and Matrix Optimization Network.
• Figure 2: DLMMPR Network.
• Figure 3: Widely Used Test Datasets in Phase Retrieval Set12.
• Figure 4: Phase Retrieval Results of Pollen Image Under Different Iteration Counts at Noise Level $\alpha = 9$ epoch/PSNR(dB).
• Figure 5: Comparison of Image Restoration Effects of Different Methods at Noise Level $\alpha =9$.