Masked autoencoder for highly compressed single-pixel imaging

The single-pixel imaging technique uses multiple patterns to modulate the entire scene and then reconstructs a two-dimensional (2-D) image from the single-pixel measurements. Inspired by the statistical redundancy of natural images that distinct regions of an image contain similar information, we report a highly compressed single-pixel imaging technique with a decreased sampling ratio. This technique superimposes an occluded mask onto modulation patterns, realizing that only the unmasked region of the scene is modulated and acquired. In this way, we can effectively decrease 75% modulation patterns experimentally. To reconstruct the entire image, we designed a highly sparse input and extrapolation network consisting of two modules: the first module reconstructs the unmasked region from one-dimensional (1-D) measurements, and the second module recovers the entire scene image by extrapolation from the neighboring unmasked region. Simulation and experimental results validate that sampling 25% of the region is enough to reconstruct the whole scene. Our technique exhibits significant improvements in peak signal-to-noise ratio (PSNR) of 1.5 dB and structural similarity index measure (SSIM) of 0.2 when compared with conventional methods at the same sampling ratios. The proposed technique can be widely applied in various resource-limited platforms and occluded scene imaging.