McLDI-INR: mask-constraint lensless dynamic imaging by dual-domain collaborative implicit neural representation

Lensless imaging, which replaces conventional optical lenses with computational techniques, enables miniaturized and lightweight imaging systems. However, inherent phase loss, manifested as twin-image artifacts and defocus noise, often necessitates multi-shot reconstruction. This dependency induces spatiotemporal coupling, inevitably compromising imaging resolution in dynamic scenarios. To address this, we propose McLDI-INR, an implicit neural representation (INR) framework that integrates temporal embedding with radial encoding for mask-constrained dynamic lensless imaging, enabling dynamic scene reconstruction from single-shot measurements. A binary encoding mask modulates the incident wavefront to build an interpretable imaging model, while joint spatial- and Fourier-domain regularization establishes a continuous spatiotemporal representation for accurate motion compensation, thereby reducing artifacts in dynamic scenes. Experiments demonstrate that our method achieves higher fidelity and resolution, offering a new paradigm for high-dimensional lensless imaging across space and time.