High turbidity underwater single-pixel imaging based on Unet++ and attention mechanism at a low sampling

Abstract

Underwater environments contain significant scattering media, characterized by complex distributions of refractive index and reflectance within. These properties cause light to deviate from its original propagation direction and decrease the signal-to-noise ratio during it passes through the scattering media. Consequently, traditional imaging techniques are difficult to achieve clear underwater imaging. Single-pixel imaging (SPI) offers higher sensitivity and the ability to be imaged in extremely low-light conditions compared to traditional imaging methods, and it provides significant advantages for underwater imaging. This paper proposes a SPI system based on Unet++ that can reconstruct high-quality images under the low sampling rate in highly turbid water environments. Numerical simulations and physical experiments demonstrate that our approach can still achieve clear image reconstruction at a sampling rate of 3.125 % under the turbidity level of 80 NTU compared with other methods. Additionally, the network exhibits strong generalization capabilities, and our approach holds the capability to reconstruct high-quality images under low sampling in high turbidity underwater conditions.