Fast high-resolution imaging combining deep learning and single-pixel imaging

Abstract

In the traditional Fourier single-pixel imaging (FSPI), compressed sampling is often used to improve the acquisition speed. However, the reconstructed image after compressed sampling often has a lower resolution and the quality is difficult to meet the imaging requirements of practical applications. To address this issue, we proposed a novel imaging method that combines deep learning and single-pixel imaging, which can reconstruct high-resolution images with only a small-scale sampling. In the training phase of the network, we attempted to incorporate the physical process of FSPI into the training process. To achieve this objective, a large number of natural images were selected to simulate Fourier single-pixel compressed sampling and reconstruction. The compressed reconstructed samples were subsequently employed for network training. In the testing phase of the network, the compressed reconstruction samples of the test dataset were input into the network for optimization. The experimental results showed that compared with traditional compressed reconstruction methods, this method effectively improved the quality of reconstructed images.