Generative adversarial network with the discriminator using measurements as an auxiliary input for single-pixel imaging

Single-pixel imaging (SPI) can realize two-dimensional imaging with a single-pixel detector without spatial resolution, and has wide application prospects in many fields because of high sensitivity and low cost. The compression reconstruction algorithm based on deep learning can improve the quality of reconstructed images. Generative adversarial network (GAN), which has excellent performance in generating images, is also gradually used in compressed sensing. However, the prior of compressively sensed measurements has not been fully utilized. Therefore, this paper proposes generative adversarial networks MAID-GAN and MAID-GAN+ with the discriminator using measurements as an auxiliary input. The image and corresponding measurements are taken as inputs of the discriminator, and the Y-shaped network structure is used to fuse the feature maps of the image domain and the measurement domain, so as to better guide the generator to generate the image close to the original image and improve the quality of the generated image. Subpixel convolution sampling is used to extract image features, and the sampling network and the reconstruction network are optimized jointly. The simulation and experimental results show that networks proposed in this paper have obvious advantages in reconstruction under low sampling rates. •The optimized sampling masks can improve the sampling efficiency.•Using measurements as an auxiliary input to the discriminator can guide the generator to generate images with more details.•Adding global features to the generator can improve the quality of reconstructed images.