Transmission-matrix-guided untrained neural network for projecting structured light through scattering media

Light projection through scattering media faces significant challenges due to random disruptions, complicating high-fidelity structured illumination essential for both everyday visualizations and industrial applications. Conventional wavefront shaping methods, like transmission matrix (TM) approaches and neural networks (NNets), suffer from pixelation artifacts and require excessive sampling, limiting their practicality. Here, we present a non-holographic projector based on a TM-guided untrained neural network (TMG-NNet) that leverages TM priors to reduce sampling demands and suppress artifacts by combining generalized physical modeling with task-specific optimization. Remarkably, this TM guidance achieves the Pearson correlation coefficient (PCC) >0.90 across diverse structured illumination tasks (MNIST digits, cell images, fringe patterns, structured beams) while reducing the required sampling rate γ = 8 — just eight times the system’s control degrees — a substantial improvement over conventional methods requiring γ ≈ 20 for the same PCC performance under non-holographic conditions.