Optimizing the quality of Fourier single-pixel imaging via generative adversarial network

Abstract

Fourier single-pixel imaging (FSI) usually achieves an enhanced imaging speed via undersampling; however, many details are lost in the reconstructed image obtained by frequency truncation due to the lack of the high-frequency part, and the image contains ringing artefacts due to the lack of an expression ability, thus reducing the image quality. To improve the quality of real-time imaging, a fast image reconstruction framework based on the Wasserstein generative adversarial network (WGAN) and gradient penalty (GP) is proposed. Under the common constraints of the resistance loss and content loss, confrontation training is first conducted between the generator and the discriminator in the framework, and then, an additional generator is connected to improve the fidelity of the reconstructed image based on the generation of the confrontation network. The model can be used to restore high-frequency details and denoise low-quality images that are undersampled. The simulation and experimental results show that the FSI using the GAN method, which achieved a compression rate of 95–98 %, is superior to conventional imaging in terms of the quality at a low sampling rate; therefore, the proposed scheme has potential practical applications.