Single Pixel Imaging Based on Enhanced Dual Generative Networks

Reconstructing high-quality image at low measurement rates remains a primary objective for single-pixel imaging (SPI). Deep learning-based compression reconstruction algorithm can improve the quality and reconstruction speed of the reconstructed image. Generative adversarial network (GAN) and diffusion model (DM), which have excellent performance in the field of image generation have achieved excellent research results in the field of single-pixel imaging reconstruction. GAN-based models often suffer from the mode collapse problem, and the diffusion model relies on multi-step iteration to obtain a fine and smooth forward diffusion process, but its lengthy sampling process limits the reconstruction efficiency. Therefore, we propose a single-pixel imaging reconstruction algorithm EDGN that combines the advantages of the two generation models. The generative adversarial network uses the measurement and noise to generate the initial image, the denoising diffusion probability model is used to optimize the image refinement, and the high-quality reconstructed image is obtained by joint training. Besides, an adaptive regularization coefficient limited range adjustment method (ARCLA) is designed to achieve fast convergence of coefficients. Simulations and experiments confirm that our proposed enhanced dual generative network method can significantly improve the quality of image reconstruction at low measurement rates.