Experimental comparison of single-pixel imaging algorithms

Liheng Bian, Jin-Li Suo, Qionghai Dai, Feng ChenarXiv 2017 · DOI · OpenReview

Abstract

Single-pixel imaging (SPI) reconstructs two-dimensional scenes from a sequence of one-dimensional light intensity measurements. While various reconstruction algorithms exist, ranging from linear correlation to non-linear compressive sensing, a unified experimental benchmark has been lacking. This paper provides a comprehensive comparison of established and newly proposed algorithms—including Conjugate Gradient Descent (CGD) and Poisson Maximum Likelihood—evaluating them on capture efficiency, computational speed, and noise robustness. Our results map the critical trade-offs necessary for selecting the optimal algorithm for specific SPI applications.

Introduction

Single-pixel imaging (SPI) is a powerful alternative to traditional array-based cameras, utilizing a single photodiode and a series of spatial light patterns to decode a scene. This architecture offers advantages in sensitivity and cost, particularly in non-visible wavebands. However, the reconstruction phase—transforming 1D intensity streams back into 2D images—presents a significant computational challenge. Historically, the field has been divided between optics-based "Ghost Imaging" (GI) and signal processing-based "Compressive Sensing" (CS). This work provides the first unified "head-to-head" comparison to guide algorithm selection based on specific task requirements, such as real-time performance or low-light accuracy.

Methods

We evaluate several classes of algorithms within a unified mathematical framework ($Ax = b$):

  1. Non-iterative (Linear Correlation): Differential Ghost Imaging (DGI), a fast but data-intensive traditional approach.
  2. Linear Iterative: Gradient Descent (GD), Alternating Projection (AP), and two newly introduced methods: Conjugate Gradient Descent (CGD) and Poisson Maximum Likelihood. CGD is specifically designed for faster convergence by ensuring search directions are conjugate to previous ones.
  3. Non-linear Iterative (Compressive Sensing): Methods utilizing image priors, specifically Sparse Representation (DCT domain) and Total Variation (TV) regularization (gradient sparsity).

The algorithms were benchmarked using both simulated data (10 standard test images) and experimental data collected from a physical SPI setup utilizing a Digital Micromirror Device (DMD) and a high-speed photodiode.

Key Results and Discussion

Our benchmarking reveals distinct trade-offs between capture efficiency and computational complexity.

Capture Efficiency vs. Sampling Ratio

Compressive sensing, particularly TV regularization, proves to be the most data-efficient. It can reconstruct high-quality images with a normalized RMSE of less than 0.03 at a sampling ratio as low as 0.5, effectively using half as many measurements as there are pixels. In contrast, linear methods like DGI require significantly higher sampling ratios to achieve comparable clarity.

Performance metrics and reconstructed images across different sampling ratios.

Scaling and Computational Complexity

While TV regularization excels at low resolutions and low sampling ratios, its computational cost increases exponentially with image size. For larger-scale reconstructions (e.g., 160x160 pixels and beyond), the newly proposed CGD and the AP methods emerge as the most practical choices. CGD demonstrates superior convergence speed, theoretically and experimentally converging within $n$ iterations, where $n$ is the pixel count.

Algorithm performance and reconstruction quality as a function of image size.

Robustness to Noise

In real-world experimental conditions characterized by high measurement noise, TV and AP demonstrate the greatest robustness. Conversely, CGD and Sparse Representation show higher sensitivity to noise, potentially limiting their utility in extremely low-light environments where the signal-to-noise ratio is poor. Additionally, DGI and AP are noted for their memory efficiency, as they do not require storing all light patterns globally.

Conclusion

This unified framework clarifies the selection criteria for SPI reconstruction. For applications requiring high data efficiency in low-light conditions, TV regularization is the optimal choice despite its computational overhead. For large-scale or real-time imaging where speed is paramount, CGD and AP provide the best balance of performance and efficiency. Future work will focus on GPU acceleration to further reduce the time penalty of high-accuracy non-linear methods.

Summary updated

Abstract

Single-pixel imaging (SPI) is a novel technique capturing 2D images using a photodiode, instead of conventional 2D array sensors. SPI owns high signal-to-noise ratio, wide spectrum range, low cost, and robustness to light scattering. Various algorithms have been proposed for SPI reconstruction, including the linear correlation methods, the alternating projection method (AP), and the compressive sensing based methods. However, there has been no comprehensive review discussing respective advantages, which is important for SPI's further applications and development. In this paper, we reviewed and compared these algorithms in a unified reconstruction framework. Besides, we proposed two other SPI algorithms including a conjugate gradient descent based method (CGD) and a Poisson maximum likelihood based method. Both simulations and experiments validate the following conclusions: to obtain comparable reconstruction accuracy, the compressive sensing based total variation regularization method (TV) requires the least measurements and consumes the least running time for small-scale reconstruction; the CGD and AP methods run fastest in large-scale cases; the TV and AP methods are the most robust to measurement noise. In a word, there are trade-offs between capture efficiency, computational complexity and robustness to noise among different SPI algorithms. We have released our source code for non-commercial use.
BibTeX
@misc{bian2017experimental,
  author = {Liheng Bian and Jin-Li Suo and Qionghai Dai and Feng Chen},
  title = {{Experimental comparison of single-pixel imaging algorithms}},
  year = {2017},
  eprint = {1707.03164},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1707.03164},
}
Figure