Computational ghost imaging based on the conditional adversarial network

Abstract A framework of computational ghost imaging based on the conditional adversarial network is proposed to efficiently implement the reconstruction of object images in this research. Although the traditional ghost imaging can recover object images with a few amount of measured intensities and corresponding speckle patterns, the visual quality of reconstructed images is seriously influenced by a lot of noise. To solve this issue, a conditional adversarial network is designed to improve the quality of the noisy images reconstructed by the traditional ghost imaging. The objective function is composed of multiple loss metrics, which avoid suffering from training instability of the network. Especially, the combination of critic loss and gradient penalty metrics guarantees that a wide variety of network architectures are used as the generator and almost no hyperparameter tuning is required in the network. Most importantly, the proposed network can be trained with simulated data without spending a long time to collect the practical training samples and considering the experimental condition. After training, the parameters of the network such as kernel weights of convolution filters are optimized from the adversarial process between the generator and the discriminator, object images can be perfectly reconstructed at a very low sampling ratio. Through the simulated and optical experiments, it is demonstrated that the proposed method has much better performance than other methods such as the traditional and compressive-sensing-based methods.