ScatResUNet: A Generalized Deep Learning Model for Imaging Through Variable-Depth Scattering Media

Scattering media presents significant challenges to the imaging process in a variety of settings, such as biomedical imaging, environmental monitoring, and remote sensing. The depths of field (DOF) play a crucial role in image quality where varying DOF can either cause focal blur or overemphasize certain regions of the image, leading to poor imaging quality. By reconstructing the affected images, deep learning (DL) has emerged as a promising solution for such challenges. This paper introduces ScatResUNet, a deep-learning model that incorporates residual blocks within the U-Net architecture to effectively reconstruct images affected by scattering media with varying scatterer axial depths. A controlled laboratory environment that simulates various scattering conditions is established to assess the efficacy of ScatResUNet. Specifically, two diffusers with adjustable axial distances are designed to generate a variety of DOF configurations in six distinct settings for two datasets, called MNIST handwritten digit and Fashion-MNIST datasets. Experimental results indicate that ScatResUNet outperforms traditional models, including U-Net and Auto encoder, in terms of critical image quality metrics. The image reconstruction accuracy of the ScatResUNet model is demonstrated by its SSIM values of 0.863 to 0.921, PSNR values of 23.388 dB to 26.561 dB, PCC values of 0.895 to 0.936, FSIM values of 0.902 to 0.953, and a JI of 0.883 to 0.932, which outperform that of the U-Net and Autoencoder. The advanced capability of ScatResUNet has the potential to significantly improve the quality of imaging in real applications of variable properties scattering environments.