Abstract
Single pixel imaging (SPI) is a novel technique being able to capture 2D images using a bucket detector with high signal-to-noise ratio, wide spectrum range and low cost. Conventional SPI projects random illumination patterns to randomly and uniformly sample the entire scene's information. Determined by the Nyquist sampling theory, SPI needs either numerous projections or high computation cost to reconstruct the target scene, especially for high-resolution cases. To address this issue, we propose an efficient single pixel imaging technique (eSPI), which instead projects sinusoidal patterns for importance sampling of the target scene's spatial spectrum in Fourier space. Specifically, utilizing the centrosymmetric conjugation and sparsity priors of natural images' spatial spectra, eSPI sequentially projects two $\pi/2$-phase-shifted sinusoidal patterns to obtain each Fourier coefficient in the most informative spatial frequency bands. eSPI can reduce requisite patterns by two orders of magnitude compared to conventional SPI, which helps a lot for fast and high-resolution SPI.
Introduction
Single-pixel imaging (SPI) offers unique advantages for capturing images in challenging environments, such as non-visible spectrums or high-noise scenarios, where traditional multi-pixel sensors are impractical or prohibitively expensive. However, the practical application of SPI has been hindered by the "speed-resolution" trade-off. Standard SPI typically requires a number of measurements proportional to the total pixel count (the Nyquist limit). While Compressive Sensing (CS) can mitigate this, it introduces significant computational overhead, often leading to long reconstruction times or memory failures for high-resolution images.
Methods
The proposed eSPI (efficient Single Pixel Imaging) framework shifts the sampling strategy from the spatial domain to the Fourier domain to exploit the inherent sparsity of natural images.
Sinusoidal Modulation and Importance Sampling
Instead of random noise, eSPI projects sinusoidal patterns that act as filters to directly extract specific Fourier coefficients. Recognizing that natural images concentrate most structural information in low-frequency components, eSPI employs an importance sampling strategy. By prioritizing the most informative frequency bands—determined by an average magnitude map of natural scenes—the system can reconstruct high-quality images with a fraction of the total possible measurements.
Exploiting Symmetry and Linear Reconstruction
To further optimize efficiency, eSPI leverages the centrosymmetric conjugation property of real-valued images in Fourier space. This mathematical symmetry allows the recovery of complex Fourier coefficients using only two phase-shifted patterns ($\pi/2$ shift), halving the requirement of previous four-step Fourier methods. Reconstruction is performed via a simple Inverse Fourier Transform (IFT), which is computationally instantaneous compared to iterative CS optimization.
Key Results
Experimental validation demonstrates that eSPI significantly outperforms traditional SPI methods in both speed and reconstruction quality at low sampling rates.
Performance Comparison
For a 128x128 image, eSPI achieves reconstruction in approximately 1 second, compared to over an hour for some CS-based methods. As shown in the comparison below, eSPI maintains high visual fidelity even at sampling ratios as low as 10%, where other methods suffer from severe noise or blocky artifacts.
Scaling to High Resolution
A key finding is that the required sampling percentage for a target quality level decreases as the total resolution increases. This scaling advantage enabled the team to reconstruct megapixel-scale images using only $10^5$ samplings, a feat that would require millions of patterns under traditional Nyquist-based SPI.
Conclusion
eSPI provides a robust solution for fast, high-resolution single-pixel imaging. By combining Fourier-domain importance sampling with symmetry-based pattern reduction, it achieves a two-order-of-magnitude reduction in required measurements. Future work will focus on adaptive sampling strategies and improving signal-to-noise ratios in high-noise environments.
Summary updated