Enabling a single pixel to "see' complex environments

Single-pixel imaging is an emerging computational imaging technique that has shown great potential in recent years for applications such as infrared detection, remote sensing, biomedical imaging, industrial inspection, and security monitoring. By projecting a series of structured illumination patterns onto a scene and recording the corresponding light intensities with a single-pixel detector, single-pixel imaging reconstructs high-dimensional images from one-dimensional temporal measurements. It provides unique advantages in spectral regions where large-scale array detectors are unavailable, such as the terahertz band, and is also well suited for wide-spectrum imaging in scattering or low-light environments.

 

However, in real-world scenarios, optical systems are inevitably affected by multiple degradations that significantly deteriorate single-pixel imaging performance. These include:

• Illumination noise from light source power fluctuations,
• Blurring caused by medium scattering or imperfections in optical components,
• Spatial jitter induced by platform or object instability, and
• Detector electronic noise and pattern-dependent noise.

 

Unlike conventional passive cameras, single-pixel imaging reconstructs images through computational algorithms, where errors in individual measurements may propagate across all pixels of the final image. As an active imaging technique, single-pixel imaging involves more noise sources and is particularly sensitive to the modeling sequence of these degradations, making its analysis more complex. Previous methods that focused on compensating single types of degradation often fail under real-world conditions involving multiple coupled degradations. While end-to-end neural networks typically rely on large, task‑specific datasets whose collection is labor‑ and resource‑intensive and lack generalization across diverse environments. Therefore, achieving high-quality single-pixel imaging in practical scenarios requires hybrid noise models that capture multiple degradations and physically consistent reconstruction methods with strong robustness and generalization.

 

In a new paper published in Light: Science & Applications, a team of scientists, led by Professor Zihan Geng from Tsinghua University, China, together with collaborators, established a physically consistent model for single-pixel imaging that incorporates multiple degradation factors to address these challenges. This model analytically describes degradations at the illumination side, detection side, and during noise propagation (Fig. 1). Based on this physics-guided noise model, synthetic signals under different degradation conditions can be generated to enhance the generalization ability of the proposed deep blind neural network. In this way, the method avoids the impractical need to precisely model and parametrize every degradation factor in real-world environments. By systematically modeling the noise processes in single-pixel imaging and performing denoising without requiring explicit knowledge of degradation parameters, the proposed method substantially improves image resolution and fidelity, offering significant theoretical and practical value.

 

To validate the method, the team built a single-pixel imaging experimental platform and conducted comparisons with other representative deep learning approaches under real-world complex degradations. Quantitative results (Fig. 2a, Fig. 2b) and color object imaging experiments (Fig. 2c) demonstrate that the proposed method consistently outperforms competing methods, highlighting its robustness and superiority in practical environments.

 

When asked about the broader significance of this work, Professor Geng commented: “Our study introduces a hybrid noise analysis and reconstruction method that requires no prior estimation of degradation parameters. This greatly enhances the robustness and imaging quality of single-pixel imaging in complex real-world conditions. We believe this technique will provide high-precision and low-cost imaging solutions for applications ranging from life sciences to industrial inspection and remote sensing.”

---

---