Physics-driven self-supervised learning for parasitic-fringe suppression in lensless holography

Parasitic fringes degrade background stability and particle-boundary contrast in lensless holographic reconstruction. In this Letter, we propose a physics-driven self-supervised framework that decouples object-related, smooth-background, and directional parasitic-fringe components within an angular-spectrum forward model. An untrained network parameterizes the object-related complex field as an implicit deep image prior, while physics consistency and component-specific spatial constraints guide the decomposition without paired data. Experiments on simulated and experimental microplastic holograms show that the method suppresses structured coherent interference, stabilizes the background, and preserves particle-boundary contrast, providing a reconstruction route for label-free particle imaging in interference-prone aqueous measurements.