Kolmogorov–Arnold network-enhanced U-Net for robust single-pixel imaging

We propose U-KAN-SPI, a U-shaped encoder-decoder framework that integrates Kolmogorov-Arnold network (KAN) structures to efficiently approximate the nonlinear mapping in single-pixel imaging (SPI). In the encoder, KANwise modules explicitly model the interactions between illumination patterns and bucket signals while suppressing noise. KAN-enhanced skip connections preserve fine structural details, and a lightweight attention module in the decoder adaptively allocates channel and spatial attention to improve robustness. Simulations and experiments demonstrate that U-KAN-SPI consistently outperforms U-Net, its attention-augmented variants, and a GAN-based reconstruction network under extremely low sampling and strong noise, yielding higher PSNR/SSIM, fewer artifacts, and sharper edges, while maintaining good generalization across different scenes. These results show that KAN structures provide an efficient module for enhancing SPI reconstruction in low-data, low-SNR regimes.