Optical Computational Imaging Through Unknown Random Diffusers in Visible Spectrum

Abstract Imaging through random scattering media is an important challenge in computational optics. Diffractive Optical Neural Networks (DONNs) have recently been demonstrated to efficiently recover images with scattering distortions using coherent light in terahertz spectrum [eLight, 2022, 2(1):4]. However, the study [Optica, 2024, 11(12):1742] has demonstrated that DONN designed for coherent light may not function optimally with low coherent sources. Consequently, the performance of DONNs in image‐based reconstruction tasks under visible and incoherent light requires further investigation. For this purpose, a three‐layer DONN is constructed to reconstruct images occluded by unknown random diffusers under coherent/incoherent visible light. The results show that the Pearson correlation coefficient (PCC) of the coherent DONNs reconstructed images can reach 0.863–0.971 for diffuser correlation lengths of (: a single neuron size of 8 µm); when of , the PCC of the incoherent diffractive optical neural networks (IC‐DONNs) reconstructed images can reach 0.861–0.899. It is found that the dynamic phase modulation mechanism introduced by randomly generated diffusers during network training enhances the adaptation of coherent DONN to spatial coherence variations of the light source. It is believed that these findings will advance the application of DONN for imaging under natural environmental conditions.