Physics-informed neural network enables high-frame-rate single-lens computational imaging

A single-lens computational imaging system combines a single lens with post-processing algorithms to achieve a lightweight design while maintaining imaging quality.However, the computational inefficiency of existing reconstruction methods often limits the achievable frame rate on edge devices, falling short of the practical requirement of 30-60 frames per second (fps).Here, we adopt a physics-informed neural network that integrates an improved Wiener deconvolution (IWD) with a compact Res-Unet variant.The simple yet effective Wiener deconvolution step reduces image blur and spatially variant degradation, thereby alleviating the workload of the subsequent network and enabling high-quality, real-time reconstruction.Simulation and experimental results demonstrate that this framework can further reduce the algorithmic complexity for a single-lens system, achieving real-time reconstruction at 40 fps for 640 480 resolution on an RK3588 system-on-chip (SoC), while maintaining a system modulation transfer function (MTF) above 0.39 at Nyquist frequency (42 lp/mm).