Learning-Based Image Reconstruction for Spatial-Variant Single-Pixel Imaging

Abstract

Improving single-pixel imaging efficiency can be achieved through spatially-variant resolution (SVR) sensing patterns, which adaptively adjust resolution to enhance image acquisition. This paper proposes a convolutional neural network (CNN) architecture specifically designed for SVR-based singlepixel imaging with compressed sensing (CS) as a more efficient and non-iterative approach to image reconstruction. The results show that the combination of SVR sensing patterns and the proposed CNN model outperforms the uniform resolution (UR) sensing patterns in terms of image reconstruction quality. Furthermore, the CNN-based approach achieves greater time efficiency compared to established methods such as ReconNet and TVAL3, thus reducing the overall computational load without compromising output image quality. These findings highlight the potential of the proposed learning-based SVR approach to effectively balance reconstruction accuracy and processing speed in single-pixel imaging. The study optimizes both the image acquisition and reconstruction process in single-pixel imaging, making it suitable for real-time applications that require rapid imaging capabilities while maintaining high image quality.

Keywords