Parallel consensus optimization for high-accuracy and robust lensless imaging.

Lensless imaging is attractive for its compact and low-cost design. However, high-accuracy and robust phase retrieval from a single pattern remains challenging and critically relies on accurate priors. Here, we propose a parallel consensus optimization termed PaCO to achieve robust and high-accuracy lensless imaging. PaCO decomposes the ill-posed problem into parallel subproblems with a distributed computational architecture. Each subproblem is handled by dedicated solvers that respectively enforce data fidelity, Fourier support, and spatial smoothness. The local solutions are aggregated into a consensus estimate. Experiments on a mask-based system demonstrate that PaCO can reconstruct objects without precise support or distance priors and accurately retrieve vortex and binary-phase-plate wavefronts, while conventional methods suffer from reconstruction instability. Statistical tests confirm that PaCO improves the success rate from 74% to 96.8% while halving the processing time. The parallel architecture of PaCO is inherently CUDA-friendly, holding strong potential for real-time processing in extreme-scale imaging workflows.