Two-Stage Training Method for High-Quality Reconstruction in Single-Pixel Imaging

A two-stage training method is proposed to enhance imaging quality and reduce reconstruction time in data-driven single-pixel imaging (SPI) under undersampling conditions. This approach leverages a deep learning algorithm to simulate single-pixel detection and image reconstruction. During the initial training stage, an L₂ regularization constraint is imposed on convolution modulation patterns to determine the optimal initial network weights. In the subsequent stage, a coupled deep learning method integrating coded-aperture design and SPI is adopted, which utilizes backpropagation of the loss function to iteratively optimize both the binarized modulation patterns and imaging network parameters. By reducing the binarization errors introduced by the dithering algorithm, this approach improves the quality of data-driven SPI. Compared with traditional deep-learning SPI methods, the proposed method significantly reduces computational complexity, resulting in accelerated image reconstruction. Experimental and simulation results demonstrate the advantages of the method, including high imaging quality, short image reconstruction time, and simplified training. For an image size of 64 × 64 pixels and 10% sampling rate, the proposed method achieves a peak signal-to-noise ratio of 23.22 dB, structural similarity index of 0.76, and image reconstruction time of approximately 2.57 × 10⁻⁴ seconds.