Single-shot color object reconstruction through scattering medium based on neural network

Abstract The reconstruction of the color object hidden behind the scattering medium is very important because the human eye is much more sensitive to color than grayscale. Traditional methods are still difficult to reconstruct an accurate image of the hidden target from one single speckle image. In this paper, a single-shot color object reconstruction technique is proposed by designing a Color Anti-scattering Convolutional Neural Network (CASNet), which is trained to output the color and structure of the hidden color target from the input of a single speckle image. The proposed technique enables us to reconstruct the target with accurate color and structure from a broadband speckle image, and the average PSNR of recovered targets with complex structure is higher than 24dB. Efficiency and accurateness are verified through experiments.