Single-pixel imaging for a phase object with a physics-driven neural network

Phase imaging technology holds significant research value in fields such as biomedicine, materials science, and precision measurement. Single-pixel imaging (SPI) for phase objects has also attracted growing attention. In existing SPI-based phase imaging approaches, phase information is typically retrieved by combining holographic techniques or complex iterative optimization algorithms. To address these limitations, we propose a self-supervised, physics-driven neural network model. By incorporating Fresnel and SPI layers into an autoencoder architecture, the network directly reconstructs the phase object from the measured intensity sequence in SPI, with network parameters updated in an end-to-end manner. Both simulation and experimental results demonstrate the effectiveness and robustness of the proposed method, showing that high-quality phase images can be reconstructed even at low sampling rates. Notably, the same SPI system can be used for both phase and amplitude objects without hardware modifications—only minor adjustments to the network structure are required. The proposed framework is also extendable to complex amplitude imaging and shows promise for applications in biomedical imaging and adaptive optics.