Imaging through scattering media with a non-symmetric U-Net

Deep learning has been widely applied in imaging through scattering media, showing advantages in penetration depth, resolution, dynamic adaptivity, et al. Due to requirements of network architecture and/or computation efficiency, the recorded speckle pattern is usually cropped or downsampled to reduce the size of the input, which often leads to severely degraded reconstruction under incoherent illumination, when target information is encoded in low-contrast speckles. In this study, we propose a non-symmetric U-Net to enhance information extraction from incoherent speckle patterns while maintaining a low cost in computation resources. The non-symmetric U-Net uses a deeper encoder to match the full-size speckle pattern and extract more feature information but a lightweight decoder to achieve target reconstruction. It has been demonstrated that the deep-contraction-shallow-expansion network structure can selectively enhance information extraction from high-frequency modulations in incoherent speckles that is overwhelmed by noise and can be easily averaged-out by downsampling. The study provides a new architectural paradigm for imaging with deep learning, which has significant research and application values in imaging fields with low coherence illuminations, such as autonomous driving detection and remote sensing.