Low-sampling and noise-robust single-pixel imaging based on the untrained attention U-Net

Abstract

The single-pixel imaging (SPI) technique illuminates the object through a series of structured light fields and detects the light intensity with a single-pixel detector (SPD). However, the detection process introduces a considerable amount of unavoidable white noise, which has a detrimental effect on the image quality and limits the applicability of SPI. In this paper, we combine the untrained attention U-Net with the SPI model to reduce noise and achieve high-quality imaging at low sampling rates. The untrained U-Net has the advantage of not requiring pre-training for better generalization. The attention mechanism can highlight the main features of the image, which greatly suppresses the noise and improves the imaging quality. Numerical simulations and experimental results demonstrate that the proposed method can effectively reduce different levels of Gaussian white noise. Furthermore, it can obtain better imaging quality than existing methods at a low sampling rate of less than 10%. This study will expand the application of SPI in complex noise environments.