High-Fidelity Computational Microscopy via Feature-Domain Phase Retrieval

Abstract

Computational microscopy enhances the space-bandwidth product and corrects aberrations for high-fidelity imaging by reconstructing complex optical wavefronts. Phase retrieval, a core technique in computational microscopy, faces challenges maintaining consistency between physical and real-world imaging formation, as physical models idealize real phenomena. The discrepancy between ideal and actual imaging formation limits the application of computational microscopy especially in non-ideal situations. Here, the feature-domain consistency for achieving high-fidelity computational microscopy is introduced. Feature-domain consistency tells that certain features, such as edges, textures, or patterns of an image, remain invariant in different image transformations, degradations, or representations. Leveraging the feature-domain consistency, Feature-Domain Phase Retrieval (FD-PR) is proposed, a framework applicable to various computational microscopy. Instead of working directly with images' pixel values, FD-PR uses image features to guide the reconstruction of optical wavefronts and takes advantage of invariance components of images against mismatches of physical models. Experimental studies, across diverse phase retrieval microscopic tasks, including coded/Fourier ptychography, inline holography, and aberration correction, demonstrate that FD-PR improves resolution by a factor of 1.5 and reduces noise levels by a factor of 2. The proposed framework can immediately benefit a wide range of computational microscopies, such as X-ray ptychography, diffraction tomography, and wavefront shaping.