Masked autoencoder for highly compressed single-pixel imaging.

Area of Science:

• Optics and Photonics
• Computational Imaging
• Signal Processing

Background:

• Single-pixel imaging (SPI) reconstructs 2D images from 1D measurements using modulation patterns.
• Natural images exhibit statistical redundancy, allowing for compressed sensing approaches.
• Conventional SPI requires extensive modulation patterns for full scene reconstruction.

Purpose of the Study:

• To develop a highly compressed single-pixel imaging technique with a reduced sampling ratio.
• To decrease the number of modulation patterns required for image acquisition.
• To enable efficient imaging in resource-limited scenarios and for occluded scenes.

Main Methods:

• Superimposing an occluded mask onto modulation patterns to acquire only unmasked scene regions.
• Designing a sparse input and extrapolation network for image reconstruction.
• Utilizing a two-module network: one for unmasked region reconstruction, another for full scene extrapolation.

Main Results:

• Experimental reduction of modulation patterns by 75%, enabling sampling of only 25% of the scene.
• Successful reconstruction of the entire scene image from limited measurements.
• Achieved 1.5 dB higher PSNR and 0.2 higher SSIM compared to conventional methods at equivalent sampling ratios.

Conclusions:

• The proposed occluded mask technique significantly compresses single-pixel imaging data acquisition.
• The sparse extrapolation network effectively reconstructs images from sparse measurements.
• This method offers a viable solution for resource-constrained imaging applications and occluded scene analysis.