Hyperspectral single-pixel imaging using untrained neural network

Single-pixel imaging (SPI) based hyperspectral imaging methods provide improved cost-effectiveness and system flexibility by eliminating the need for complex mechanical scanning that is commonly used in conventional HSI setups. However, traditional algorithms used in SPI still have some shortcomings, including a large number of iterations and poor image quality at low sampling rates. Furthermore, in deep learning based SPI approaches, the quality of the reconstructed images is heavily dependent on the training dataset. The training process for neural networks is often computationally intensive and time-consuming, posing a major bottleneck for resource-constrained applications. To address these challenges, we introduce a novel single-pixel hyperspectral imaging (SP-HSI) method that uses a minimized multi-channel sensor (16-channel) as the detectors, and employs a physics-driven neural network for high-fidelity image reconstruction without pre-training. The framework of the network comprises a physical model of SP-HSI and a U-Net model. By simulation, we demonstrate the reconstruction of 32-channel hyperspectral images. This work helps to develop next-generation hyperspectral imaging technologies for scalable and portable applications.