Abstract
Lensless cameras based on thin diffusers offer a compact alternative to conventional refractive imaging but rely on computational reconstruction, since the diffuser’s point spread function (PSF) globally multiplexes every scene point across the sensor. Here we develop a Fourier neural operator (FNO) framework for this reconstruction task. Because a linear shift-invariant forward model reduces to a pointwise multiplication in Fourier space, the spectral-domain kernel of an FNO layer is structurally aligned with the diffuser inverse problem: its learned weights act globally in frequency, matching a PSF that spans the sensor, and are indexed by physical frequency rather than by a fixed pixel grid, so the operator is by construction independent of the measurement discretization. We characterize the framework along the two axes that matter for deployment—reconstruction quality at a fixed parameter budget relative to convolutional and physics-informed baselines, and behavior under changes in measurement discretization—and introduce a fixed-capacity receptive-field analysis that separates the contribution of global spatial support from that of the spectral parameterization. This structural view indicates that the reconstruction problem, a global convolution naturally inverted in the spectral domain, is a useful guide for the choice of network architecture, and that neural operators provide a practical path toward resolution-agnostic lensless imaging, with the framework extending to other lensless modalities governed by a global PSF.
Funding Information
  • Türkiye Bilimsel ve Teknolojik Araştırma Kurumu (122C150)

This publication has 25 references indexed in Scilit: