Single-pixel imaging with TCAO and untrained neural network

Single-pixel imaging (SPI) is a rapidly evolving computational imaging technique that reconstructs scenes by correlating modulation patterns with measurements captured by a single-pixel detector. Recent advances suggest that integrating model-driven deep learning can significantly enhance the reconstruction quality and robustness of SPI. However, current model-driven SPI methods often rely on ghost imaging (DGI) with random speckles as network input, requiring deeper reconstruction networks to extract effective features, which increases the computational cost. Additionally, random speckles can cause important image details to be obscured by noise at lower sampling rates, making it challenging for the network to produce satisfactory reconstructions. To overcome these limitations, we propose a model-driven SPI method that utilizes an optimized sorting of the Hadamard matrix, termed Total Change Ascending Order (TCAO), as the modulation mask, coupled with an untrained convolutional neural network (CNN) for reconstruction. TCAO is designed to more effectively extract information from scenes at lower sampling rates. The core innovation is integrating deep learning principles across the entire imaging process, assigning more feature extraction tasks to the modulation stage. We refer to this approach as Deep Learning-Based Single-Pixel Imaging with Efficient Sampling (DLES). Simulation results show that DLES allows the network to focus on enhancing reconstruction performance, yielding superior results at lower and even extremely low sampling rates. This method provides a novel approach to simplifying model-driven neural networks while improving the efficiency and quality of single-pixel imaging.