Compressive hyperspectral phasor imaging with single-pixel detection for spectral tasks

Abstract: Spectral vision task plays a pivotal role in extracting discriminative spectral-spatial features from high-dimensional data, enabling fine-grained identification beyond human vision. Traditional methods usually involve first collecting rich spectral-spatial information and then using complex algorithms to digitally process it into scene classification and recognition. However, the complexity of processing massive three-dimensional (3D) hyperspectral datasets poses challenges for algorithms. Here, we demonstrate a compressive Hyperspectral Phasor Imaging with Single-pixel detection (HyPIS) that leverages highly compressed spatial-spectral data to achieve spectral task. Two optical encoders are used for wavelength-dependent sine- and cosine-encoding that transforms spectral signals into a two-dimensional (2D) phasor plot. By applying spatial-temporal illumination patterns, a single-pixel detector is enough to reconstruct the phasor image of the object. This allows to directly generate pixel-wise spectral task, bypassing 3D hyperspectral data. Our experiments show that HyPIS can perform real-time classification and recognition tasks of different scenes, reducing the required amount of data by two orders of magnitude, and it can still accurately classify under low light and uneven lighting conditions. This work develops a completely new spectral technology that enables spectral tasks to be performed without obtaining high-resolution hyperspectral datasets, holding promise for spectral applications in mobile devices, robotics, and satellite technologies.

• The paper presents HyPIS, a dual-compression system using customized sine and cosine encoders paired with single-pixel imaging for efficient hyperspectral data reduction.
• It achieves up to a 100-fold data volume reduction while enabling direct spectral classification through a phasor mapping technique without conventional machine learning pipelines.
• The system demonstrates robust spectral discrimination under low and uneven illumination, supporting real-time performance for portable and embedded spectral applications.

Compressive Hyperspectral Phasor Imaging with Single-Pixel Detection: A Technical Appraisal

Introduction and Motivation

This work introduces the HyPIS (Hyperspectral Phasor Imaging with Single-pixel detection) architecture, targeting data-efficient hyperspectral classification and recognition. Conventional hyperspectral imaging demands an acquisition-storage-computation workflow, typically constrained by the handling of massive spatiotemporal-spectral datacubes. Existing dimensionality reduction paradigms (e.g., linear unmixing, PCA, or nonlinear manifold learning) only decouple redundancy post-acquisition and provide limited relief against data transfer and storage bottlenecks. Hardware-embedded alternatives such as spectral kernel machines integrate inference with sensing but remain limited by the training scope, impeding their applicability in out-of-distribution or unpredictable environments.

HyPIS addresses these limitations by combining optical domain phasor encoding and computational single-pixel imaging (SPI), realizing direct task-oriented classification without reconstruction of high-dimensional datacubes or reliance on pre-trained models. Through 2D compression along both spectral and spatial axes, HyPIS dramatically reduces acquisition complexity and downstream computational burden.

Methodology: Dual-domain Compression and Phasor Encoding

HyPIS synergistically integrates spatial SPI and wavelength-dependent optical encoding. Spatial compression is achieved via structured illumination (orthogonal DMD-generated patterns), while the spectral compression leverages dual optical encoders (sinusoidal and cosinusoidal transmission profiles).

For each illumination pattern, HyPIS records three signals through different optical paths: (i) unencoded reference, (ii) sine-encoded, and (iii) cosine-encoded. After data acquisition, a variant of compressed sensing-based image reconstruction yields spatial maps for each encoding. Subsequently, phasor analysis projects the spectral signature of each pixel to the $(G, S)$ 2D phasor plane, where

$$G(x, y) = 2 \frac{O_{\text{cos}}(x, y) - I_{\text{cos,min}}}{I_{\text{cos,max}} - I_{\text{cos,min}}} - 1,$$

$$S(x, y) = 2 \frac{O_{\text{sin}}(x, y) - I_{\text{sin,min}}}{I_{\text{sin,max}} - I_{\text{sin,min}}} - 1.$$

The position in the phasor plane encodes the underlying spectral information, enabling direct clustering and object classification.

This framework circumvents the need for 3D datacube acquisition, algorithmic spectral unmixing, or supervised model training, facilitating low-latency, hardware-integrated spectral inference.

Numerical and Experimental Results

Simulations on standard hyperspectral datasets (e.g., CAVE) demonstrate that HyPIS reduces required storage by one to two orders of magnitude compared to conventional methods—achieving equivalence at 1/15.5 of raw data volume—while maintaining scene classification accuracy.

Bench-scale experiments validate the spatial-spectral discrimination achievable by HyPIS. In static transmission mask experiments, phasor clustering in $(G,S)$ space matches conventional spectrometer-based phasor results and correctly classifies all spatial regions, including objects with highly similar apparent colors (e.g., metameric pairs). The experimental verification demonstrates that storage requirements for HyPIS are 1/500 that of traditional approaches for full-color image reconstruction or classification tasks.

Under dynamic conditions (rotating object mask, 2.68 fps at $64 \times 64$ resolution), HyPIS stably classifies objects across frames even under extreme nonuniform illumination and low SNR, leveraging its insensitivity to global brightness and spatially varying signal intensity. Notably, HyPIS differentiates metameric objects with comparable visual color but distinct spectral characteristics—a critical advantage for applications in spectral imaging where color-based methods fail.

Practical Implementation and Robustness

The HyPIS optical architecture utilizes cost-effective, commercially available DMDs, single-pixel detectors, and custom-fabricated transmissive optical encoders. Experimental data validate that deviations from ideal sinusoidal/cosinusoidal encoder transmission profiles introduce only minor geometric deformation in the phasor representation, correctable via calibration. The robustness under varying illumination intensity, detector gain, and environmental inhomogeneity is quantitatively supported in control studies.

Further, proof-of-concept deployment in reflective mode and demonstration on plant targets (Anthurium leaves and stems) confirm that HyPIS enables spatial recognition tasks using only a pre-registered phasor reference database. Such capability is central for field applications where online recalibration or retraining is impractical.

Theoretical and Practical Implications

The HyPIS approach substantiates an alternative paradigm for spectral machine vision: direct task inference at the sensor level, decoupled from traditional data pipeline constraints. Immediate implications include substantial reduction in the bandwidth and memory requirements for mobile, wearable, and remote imaging platforms, and direct suitability for integration with UAVs, robotics, or satellites.

The method's extension to wavebands outside the visible (NIR, MIR, THz) is theoretically justified by the hardware-agnostic nature of the optical encoding, limited only by material and detector constraints. The framework also admits straightforward fusion with DNN inference modules for higher-level semantic tasks under bandwidth- or acquisition-limited regimes.

For future development, the engineering of structured metasurfaces or low-dimensional material encoders for on-chip phasor encoding, integration with 2D material photodetectors, and combination with physics-informed neural networks represent direct research trajectories. These extensions promise further reduction in system size, cost, and energy consumption, and broader applicability in spectral imaging tasks where speed, size, or energy constraints preclude use of conventional imagers.

Conclusion

HyPIS is established as a dual-domain compressed hyperspectral imaging architecture that integrates optical phasor encoding and single-pixel detection to enable real-time, low-memory, and hardware-efficient spectral task performance. By eliminating the dependency on high-dimensional data acquisition and post-acquisition computational pipelines, HyPIS delivers strong spatial-spectral discrimination—even in challenging visibility or spectral proximity scenarios—while being robust and adaptable to practical constraints. Anticipated future directions include extension to broad spectral bands, integration with intelligent algorithms, and full miniaturization for pervasive spectral sensing platforms.