Ghost Imaging by Single-Layer Neural Network Based on Forward Physical Model

Ghost imaging (GI), due to its unique imaging principle, has great potential for applications in hyperspectral imaging, 3D imaging, and remote sensing observations. However, the conflict between imaging quality and measurement times has hindered the practical application of GI, becoming a focus of many studies. In order to reconstruct higher quality images with fewer measurements, this paper proposes an image reconstruction algorithm that uses a single-layer neural network to fit the forward physical model of GI. The network takes a series of projected light fields as input and is constrained by the physical model, which enables the output to approach the true detection values, thus gradually making the reconstructed image closer to the ground truth. Since the inputs and labels are known data, the network does not need to be pre-trained. Both simulations and experiments have verified the effectiveness of this approach, while its convergence, flexibility, robustness and reconstruction time have also been discussed. The results indicate that this method can recover high-quality images at low sampling ratios, and can perfectly reconstruct objects from undersampled data. The imaging quality and efficiency surpass those of some traditional GI techniques. This not only promotes the practical application of GI but also provides the possibility of real-time imaging using untrained networks.