• Proposes a diffusion model-based Fourier single-pixel non-line-of-sight (NLOS) imaging method. • Achieves high-quality NLOS imaging under conditions involving transmission through six A4 paper layers and extremely low sampling rates of 3 %. • Enables superior NLOS performance even in scenarios with strong obstructions, multi-layer scattering, or challenging acquisition conditions. • The method can expand the application potential of Fourier single-pixel imaging in complex environments. Non-line-of-sight (NLOS) imaging is aimed at the reconstruction of objects obscured from direct view. Recent advancements in techniques, including Time-of-flight (ToF), ghost imaging, and imaging through scattering media, have significantly advanced NLOS imaging. However, challenges including poor real-time performance, high sampling rates, and limited adaptability to complex environments persist. Fourier single-pixel imaging (FSPI) emerges as a highly promising technique for addressing NLOS imaging challenges. Nevertheless, traditional FSPI suffers from severe loss of high-frequency information at extremely low sampling rates. Deep learning approaches, including deep convolutional generative adversarial network (DCGAN) and diffusion models, offer new avenues for high-quality reconstruction under low sampling conditions. Here, a novel NLOS FSPI imaging method based on a diffusion model was proposed. During data acquisition, multiple layers of A4 paper are used to simulate line-of-sight occlusion, and a single-pixel detector captures the transmitted light signals through these layers to construct a low-frequency spectrum. This spectrum is then fed into a pre-trained diffusion model for iterative reconstruction. By introducing Gaussian noise to perturb the data distribution during training, the model learns prior information, effectively recovering high-frequency details. In the reconstruction process, the low-frequency Fourier spectrum serves as a consistency constraint, iteratively optimized to yield high-quality images. Under conditions involving transmission through multiple A4 paper layers and extremely low sampling rates (e.g., 3%), the proposed method significantly enhances the reconstruction quality of NLOS target images compared to traditional FSPI approaches. This method provides a novel approach for NLOS FSPI applications under low sampling rates, demonstrating substantial practical potential.
Fu et al. (Thu,) studied this question.