Transformer-Based Under-sampled Single-Pixel Imaging

Single-pixel imaging, as an innovative imaging technique, has attracted much attention during the last decades. However, it is still a challenging task for single-pixel imaging to reconstruct high-quality images with fewer measurements. Recently, deep learning techniques have shown great potential in single-pixel imaging especially for under-sampling cases. Despite outperforming traditional model-based methods, the existing deep learning-based methods usually utilize fully convolutional networks to model the imaging process which have limitations in long-range dependencies capturing, leading to limited reconstruction performance. In this paper, we present a transformer-based single-pixel imaging method to realize high-quality image reconstruction in undersampled situation. By taking advantage of self-attention mechanism, the proposed method is good at modeling the imaging process and directly reconstructs high-quality images from the measured one-dimensional light intensity sequence. Numerical simulations and real optical experiments demonstrate that the proposed method outperforms the state-of-the-art single-pixel imaging methods in terms of reconstruction performance and noise robustness.