Imaging reconstruction through long-range scattering media by using deep learning

The realization of high-resolution imaging of images through scattering media has always been an important problem to be solved in the field. In this paper, our purpose here is to create a new framework that can realize the imaging through long-range scattering media. To do so, we establish a long-range scattering medium model, and use the model to generate simulated speckle pattern. In particular, we are designing a new neural network that is able to learn the statistical information found in the pattern of speckle intensity. The simulated speckle data were used as train sets for the neural network, and the learning rate of the SGD was 0.001, so that the model converged, which had good effects in the aspects of recovery time, imaging quality, mobility, convergence rate and so on. The peak signal to noise ratio (PSNR), Pearson correlation coefficient (PCC), structural similarity (SSIM) and other indexes were used to evaluate the performance of the convolution neural network in restoring images. Our neural network has achieved good results under this evaluation index from those results. PSNR value is 16.939, SSIM value is 0.842, and PCC value is 0.884, indicating that our new neural network model can realize long-range scattering media imaging and improve the imaging quality of scattering imaging.