Lensless Phase Retrieval With Regularization By Blind Noise Map Estimation and Denoising

Abstract

This paper addresses the challenge of regularization in lensless single-shot phase retrieval (PR) by noise suppression. Due to the unique aspects of the PR algorithm, the noise is spatially correlated with a non-stationary level and distribution, which complicates PR’s reconstruction and convergence. To address this problem, We propose an algorithm for noise suppression, which utilizes the PIXPNet network for the initial estimation of noise parameters and prefiltering noise associated with the heavy tails of the noise distribution. Subsequently, the DRUNet network is applied within frequency sub-bands to suppress the noise meticulously. Our findings reveal that the proposed regularization, operating in a fully blind mode, outperforms our previous PR algorithm by achieving more effective noise suppression, enlarged field of view, and enhanced accuracy in estimating the height map of the object.