Machine Learning Enables 96% Accurate Single-Pixel Detection Of Structured Light With Speckle-Learned Recognition

Structured light, with its unique spatial properties, holds immense promise for advancements in communications, precision measurement, and sensing technologies. Purnesh Singh Badavath and Vijay Kumar, from the National Institute of Technology Warangal, present a significant step forward in how these complex light beams are identified and analysed. Their research demonstrates a new method for recognising structured light using only a single detector, moving beyond the need for complex optical setups and high-resolution imaging. By mapping the information contained within structured light onto simple, one-dimensional speckle patterns captured in both space and time, the team achieves remarkably high accuracy, exceeding 94% and 96% respectively, while also proving resilience to environmental disturbances and ambiguities, paving the way for practical, scalable applications in optical systems.

Sensing typically relies on methods demanding complex alignments and high-resolution imaging. Speckle-learned recognition, or SLR, presents a powerful alternative, exploiting the spatio-temporal speckle fields generated by light-diffuser interactions. This work builds upon earlier research, expanding the capabilities of speckle-based recognition systems and offering a potentially simpler and more robust method for structured light detection.

Diffuser-Based Structured Light Recognition Achieved

This research demonstrates significant advances in structured light recognition, moving beyond traditional methods that require precise alignment and high-resolution imaging. Scientists have successfully developed and tested methods for identifying complex light beams, including Laguerre-Gaussian and Hermite-Gaussian modes, by analyzing the patterns created when light interacts with a diffuser. Crucially, this work achieves high accuracy, exceeding 94% in spatial and 96% in temporal domains, while using substantially less data than conventional approaches. The team achieved this by mapping two-dimensional information about the light beams onto one-dimensional speckle patterns, captured either in space or time.

Spatial methods reduce data requirements and computational load, making them suitable for resource-constrained applications. Temporal methods, which utilize a single-pixel detector to capture fluctuations over time, prove particularly robust to turbulence and can even distinguish between beams that appear identical using other techniques. The research highlights the potential for scalable, low-power, and energy-efficient optical systems for communication and sensing. Future work will focus on applying these techniques to areas such as high-speed free-space optical links, biomedical diagnostics, material inspection, and the development of sustainable photonic technologies for remote communication and on-chip devices. This work establishes a new paradigm for structured light recognition, offering a fast, robust, and sustainable alternative to existing methods.

Speckle Patterns Identify Structured Light Beams

Structured light (SL) beams, such as Laguerre-Gaussian, Hermite-Gaussian, and perfect vortex beams, possess engineered amplitude, phase, or polarisation distributions and offer new degrees of freedom for applications including optical trapping, imaging, metrology, and high-capacity communication systems. Reliable recognition of these beams is therefore essential. Conventional methods, based on diffraction, interference, or modal decomposition, can be limited, motivating the exploration of machine learning approaches. Researchers have developed speckle-based structured light recognition techniques, progressing from two-dimensional (2D) image analysis to more efficient one-dimensional (1D) approaches.

A 1D speckle-learned method maps 2D speckle images onto 1D line arrays, retaining essential mode-dependent features while significantly reducing data size, computational cost, and training time compared to 2D methods. Further advancements have led to temporal zero-dimensional speckle-learned recognition, implemented by mapping temporal speckle sequences with a single-pixel detector. This approach robustly recognises intensity degenerate structured light beams, maintaining accuracies above 95% even under severe turbulence. By mapping spatial information into temporal sequences, the method eliminates sensitivities to anisotropy and asymmetry observed in spatial schemes.

Both spatial 1D and temporal 0D methods are scalable, energy-efficient, and computationally lightweight, enabling faster training and recognition with minimal hardware requirements. These techniques offer sustainable alternatives to traditional camera-based recognition, lowering power consumption and system complexity while preserving high fidelity in beam classification. Looking ahead, these methods have potential in high-capacity free-space optical links, biomedical diagnostics, material inspection, and metrology. Their energy efficiency and low computational cost align with the need for sustainable photonic technologies, suitable for satellite links, remote communication, and on-chip photonic devices.

Speckle Learning Recognises Structured Light Beams

Scientists have achieved a breakthrough in structured light recognition by developing innovative speckle-learned recognition (SLR) schemes that operate across both spatial and temporal domains. This work builds upon earlier research and significantly advances the ability to identify complex light beams, including Laguerre-Gaussian (LG) and Hermite-Gaussian (HG) modes, with increased efficiency and robustness. Experiments demonstrate that by mapping two-dimensional speckle patterns onto one-dimensional spatial arrays, the team successfully recognised structured light beams with accuracy exceeding 94%, even while using only a fraction of the original data. Further refinement of this technique involved mapping the spatial information of structured light onto temporal speckle sequences recorded with a single-pixel detector.

This temporal approach achieved accuracies exceeding 96% across various structured light families, demonstrating resilience to turbulence and even when beams exhibit modal degeneracy. Specifically, a support vector machine model trained on these temporal sequences achieved 86. 7% and 92. 5% accuracies for LG and HG beams respectively. Investigations into the influence of speckle grain size relative to detector size revealed that classification accuracy improves as the detector size increases relative to the speckle grain size, with a critical ratio of DS/SS ≥ 1.

13 required for reliable classification. Importantly, the temporal approach demonstrated strong robustness even under severe turbulence, maintaining classification accuracies above 95% with turbulence levels up to Cn2 = 1 × 10−12m −2/3, a significant improvement over spatial methods which experienced accuracy reductions under similar conditions. These advancements deliver scalable, energy-efficient, and computationally lightweight recognition architectures, reducing data size and training time compared to traditional two-dimensional camera-based systems. The team’s work establishes a pathway for practical implementations in free-space optical communication and real-time sensing, offering a sustainable alternative while preserving high fidelity in beam classification.

👉 More information
🗞 Machine learning meets Singular Optics II: Single-pixel Detection of Structured Light
🧠 ArXiv: https://arxiv.org/abs/2509.16946