Efficient Kronecker-based deep unfolding network for single-pixel imaging with hardware implementation

Single-pixel imaging (SPI) is a novel computational imaging technique. Recently, deep learning has achieved remarkable improvements in SPI reconstruction in both quality and speed. However, the rapidly increasing computational costs at large image sizes remain a key obstacle to practical SPI. In this work, we propose the K-CPP Net, a gradient-descent-based deep unfolding network that couples a Kronecker-based SPI model with a Chambolle-Pock-inspired primal-dual iterative optimization algorithm for efficient SPI reconstruction. Specifically, we avoid explicit vectorization of the large sensing matrix by representing the matrix as the Kronecker product of two learnable subsampling matrices, and design a lightweight denoising module based on residual convolutional blocks and efficient channel-wise self-attention. This method substantially reduces the computational cost, enabling full-image SPI training and inference while effectively suppressing distortion and blur and preserving fine structural details. Moreover, we implement K-CPP Net on an FPGA via an optimized accelerator architecture, providing a reference design for real-world SPI applications. Experimental results demonstrate that the proposed method achieves superior reconstruction quality while offering markedly improved computational efficiency.