Optimizing Under-Sampling in Fourier Single-Pixel imaging using GANs and attention mechanisms

Single-Pixel Imaging (SPI) is an emerging imaging technique with unique advantages, but Fourier Single-Pixel Imaging (FSPI) still faces challenges in improving sampling efficiency and reconstruction quality. The generation of modulation patterns in FSPI typically requires the creation of a linear sampling space, and consequently, the complexity of the sampling space directly impacts the imaging efficiency of FSPI. This paper proposes an FSPI under-sampling optimization method based on Generative Adversarial Networks (GANs) and attention mechanisms, aiming to dynamically generate optimized sampling masks and enhance reconstruction efficiency. Unlike existing end-to-end approaches, this method employs a GANs and Monte Carlo model to directly generate sampling masks instead of images, providing greater flexibility and optimization potential. The generator adopts U-net with ResBlock structure to enhance gradient propagation. This method can capture details and high-frequency information in the image, consequently improving the reconstruction quality. Moreover, through adversarial training, the generator can produce diverse and realistic sampling patterns, thus enhancing the generalization ability of the model. The effectiveness and superiority of this method have been validated by various experiments. In the optical experimental part, we designed a scattering media SPI system to validate the proposed method and compared and analyzed the optimal detector gains. Notably, our method demonstrated excellent adaptability in scattering media conditions, achieving high-quality image reconstruction even in the presence of A4 paper and a water tank as scattering media. This research provides new insights and solutions for the application of deep learning in FSPI under-sampling optimization, particularly in complex imaging environments.