Fourier computational ghost imaging using spectral sparsity and conjugation priors

Liheng Bian, Jinli Suo, Xuemei Hu, Feng Chen, Qionghai Dai

Abstract

Computational ghost imaging (CGI) retrieves a target scene from numerous random illumination patterns and corresponding single pixel measurements. Theoretically, these random patterns sample random combinations of the Fourier coefficients of the scene's spatial spectrum in an indiscriminative way, and neglect their intrinsic nonuniform importance. Utilizing the sparsity and conjugation priors of natural images' spatial spectra, this letter proposes a new pattern illuminating strategy termed Fourier computational ghost imaging (FCGI), for highly efficient single pixel imaging. Specifically, FCGI sequentially uses two sinusoidal patterns to sample each Fourier coefficient, instead of their random combinations, in the statistically most informative spectrum band. Benefiting from the new illumination patterns and importance sampling strategy, our approach is able to reduce the requisite number of projected patterns by two orders of magnitude, compared to conventional CGI.

Introduction: The Speed Limit of Single-Pixel Imaging

Single-pixel imaging (SPI) offers a powerful alternative to traditional array-based cameras, particularly for imaging in non-visible spectra or through challenging media. By using a single "bucket" detector and a series of structured light patterns, 2D scenes can be reconstructed from 1D measurements. However, standard SPI typically relies on random illumination patterns, which often require a massive number of measurements—sometimes exceeding the total pixel count—to achieve high resolution. While Compressive Sensing (CS) can mitigate this, the resulting computational overhead is immense, often requiring hours of processing and significant memory, which hinders real-time applications.

Methodology: Efficient Single-Pixel Imaging (eSPI)

The proposed efficient Single Pixel Imaging (eSPI) framework shifts the imaging problem from the spatial domain to the Fourier domain. This transition enables two key optimizations:

1. Importance Sampling

Natural images are not random; their energy is typically concentrated in low-frequency bands. By analyzing the average spectrum of natural scenes, the authors identified a "diamond-shaped" region in the Fourier domain that contains the most critical information. Rather than sampling the entire spectrum uniformly, eSPI prioritizes these high-importance frequencies.

Spectral importance sampling in natural images.

2. The Two-Step Sinusoidal Strategy

Traditional Fourier-based SPI requires four phase-shifted sinusoidal patterns to resolve a single complex Fourier coefficient. eSPI exploits the fact that natural scenes are real-valued, meaning their Fourier spectra are centrosymmetric. This symmetry allows the researchers to reconstruct the scene using only two $\pi/2$-phase-shifted patterns per frequency, effectively halving the required measurements compared to previous Fourier methods. Reconstruction is performed via an Inverse Fast Fourier Transform (IFFT), which is computationally instantaneous compared to iterative optimization.

Results and Performance

The eSPI method was validated through both simulations and a physical prototype using a Digital Micromirror Device (DMD).

Comparative Performance

When compared to conventional random-pattern SPI (using both linear and CS reconstruction), eSPI demonstrated superior image quality at significantly lower sampling ratios. For instance, at a 10% coverage ratio, eSPI produced sharp reconstructions while random-pattern methods remained noisy or blurred.

Comparison of eSPI with conventional random-pattern methods at 10% and 30% sampling ratios.

Speed and Scalability

The computational efficiency of eSPI is its most striking feature. In a $128 \times 128$ pixel simulation, eSPI achieved reconstruction in just 1 second, whereas CS-based methods took over an hour to reach comparable quality. Furthermore, as image resolution increases, the relative sparsity of the Fourier spectrum allows eSPI to scale efficiently; a megapixel image can be reconstructed from a tiny fraction of the total possible samples without the memory crashes common in CS algorithms.

Conclusion

By moving from random spatial patterns to targeted Fourier-domain sampling, eSPI overcomes the primary bottlenecks of single-pixel imaging. The combination of importance sampling and spectral symmetry exploitation reduces measurement requirements by two orders of magnitude and enables near-instantaneous reconstruction. Future work will focus on adaptive sampling strategies and mitigating noise sensitivity in real-world environments.

Summary updated

Abstract

Computational ghost imaging (CGI) retrieves a target scene from numerous random illumination patterns and corresponding single pixel measurements. Theoretically, these random patterns sample random combinations of the Fourier coefficients of the scene's spatial spectrum in an indiscriminative way, and neglect their intrinsic nonuniform importance. Utilizing the sparsity and conjugation priors of natural images' spatial spectra, this letter proposes a new pattern illuminating strategy termed Fourier computational ghost imaging (FCGI), for highly efficient single pixel imaging. Specifically, FCGI sequentially uses two sinusoidal patterns to sample each Fourier coefficient, instead of their random combinations, in the statistically most informative spectrum band. Benefiting from the new illumination patterns and importance sampling strategy, our approach is able to reduce the requisite number of projected patterns by two orders of magnitude, compared to conventional CGI.
BibTeX
@article{bian2016fourier,
  author = {Liheng Bian and Jinli Suo and Xuemei Hu and Feng Chen and Qionghai Dai},
  title = {{Fourier computational ghost imaging using spectral sparsity and conjugation priors}},
  journal = {Journal of Optics},
  year = {2016},
  doi = {10.1088/2040-8978/18/8/085704},
}

Loading related content...

Figure