Color-Encoded Illumination Techniques

• Color-encoded illumination is a method that systematically varies light spectra to extract, transmit, and visualize hidden scene and material properties.
• It employs techniques such as temporal strobing, spectral optimization, and multiplexed coding to enhance applications in high-speed imaging, multispectral sensing, and visible light communication.
• Precise calibration and advanced algorithms, including CNN-based models and convex optimization, are critical for accurately mapping designed illumination parameters to sensor responses.

Color-encoded illumination encompasses a diverse set of physical, algorithmic, and computational techniques wherein the spectrum, chromaticity, or modulation of illumination is systematically varied or encoded to extract, transmit, modify, or visualize information in ways not accessible via conventional broadband or white lighting. The concept drives state-of-the-art advances in computational color constancy, high-speed volumetric imaging, multispectral sensing, advanced communication systems, and visualization. Central to all approaches is the rigorous mapping from designed illumination spectra and temporal/color coding to measurable responses—whether in camera raw space, spatial detectors, or human-perceptual color—enabling robust inference or control of scene properties, material appearance, or embedded signals.

1. Principles and Mathematical Foundations

At its core, color-encoded illumination leverages the interdependence of spectral illumination, material reflectance, and detector (or human) spectral sensitivity. Mathematically, the basic colorimetric model stipulates that the measured color at a pixel or sensor is given by

$$\mathbf{v} = \mathbf{C} \cdot \mathbf{R} \cdot \mathbf{s}$$

where $\mathbf{s}$ is a (possibly time- or channel-coded) spectral power distribution (SPD) of the light source, $\mathbf{R}$ is the material’s spectral reflectance (possibly spatially varying or unknown), and $\mathbf{C}$ comprises color-matching or sensor-response functions. Encoding is achieved by systematically manipulating $\mathbf{s}$—either as a function of time, spectrum, spatial mask, or jointly with modulation waveforms—and then mathematically or algorithmically decoding the resultant measurements to infer scene properties or transmit information.

Specific instances include:

• Sampling illuminant chromaticities along the Planckian locus for color constancy, by mapping target illuminants via calibrated matrices into camera-native raw color space and encoding the chromatic trajectory (Kim et al., 10 Apr 2025).
• Rapid sequential color strobing to map time-resolved scene evolution onto color mixtures captured during a single camera exposure (Novikov et al., 29 Apr 2026).
• Spectral optimization for metameric manipulation whereby the SPD is engineered to control perceived object color under a white-point constraint (Yamaguchi et al., 2024).
• Multiplexed color coding and compressive sensing for single-pixel or low-data-rate multispectral imaging (Huang et al., 2017).
• Channel-multiplexed visible light communications (VLC) by exploiting color, frequency, and DC level diversity (Gao et al., 2015, Cinemre et al., 13 Apr 2025).

2. Color-Encoding Methodologies Across Domains

A range of methodologies implement color-encoded illumination, each adapting fundamental principles for specific scientific, computational, or engineering goals.

Computational Color Constancy

Modern cross-camera color constancy networks (e.g., CCMNet) exploit calibrated color correction matrices (CCMs) specific to each camera to encode how the locus of ideal white point colors (the Planckian locus) is mapped into the camera's raw RGB space. For correlated color temperatures $t\in\{2500K,\dots,7500K\}$, each corresponding CIE XYZ vector $X_t$ is mapped via

$L_t = CCM_t \, X_t$

Log-chroma coordinates $(u,v)$ are computed as $u=\log(L_G/L_R)$, $\mathbf{s}$0, and the locus is aggregated spatially into a histogram. A compact embedding (Camera Fingerprint Embedding, CFE) is generated by passing this histogram through a CNN, providing robust camera-specific guidance for cross-domain color constancy estimation (Kim et al., 10 Apr 2025).

High-Speed Scene Capture via Temporal Color Coding

In high-speed 3D scene reconstruction, temporal information is mapped to color encoding using rapid sequences of color strobes during a single low-speed camera exposure. Each subframe $\mathbf{s}$1 employs a triplet $\mathbf{s}$2 to define the instantaneous RGB illumination:

$\mathbf{s}$3

The resulting camera measurement is a linear mixture:

$\mathbf{s}$4

where $\mathbf{s}$5 are the latent high-speed grayscale frames. Decoding high-speed volumetric content involves disentangling these mixtures via both calibration and inverse modeling using dynamic Gaussian Splatting (Novikov et al., 29 Apr 2026).

Spectral Control and Metamerism

Targeted color appearance manipulation by illumination employs SPD optimization to achieve specific colorimetric transformations for select materials under a white-point or CRI constraint. The optimization minimization is expressed as

$\mathbf{s}$6

subject to white-point and CRI constraints, and parameterized on an $\mathbf{s}$7-channel LED basis $\mathbf{s}$8 (Yamaguchi et al., 2024).

Multiplexed Computational Imaging and Communication

• In multispectral ghost imaging, three orthogonal binary masks encode R, G, B spatial channels into a single multiplexed measurement. The overall linear measurement model is $\mathbf{s}$9, with $\mathbf{R}$0 comprising color- and mask-multiplexed spatial structure (Huang et al., 2017).
• In advanced VLC, joint color-frequency and DC-biased constellations maximize SNR under physical constraints, leveraging high-dimensional sphere-packing (Gao et al., 2015); the DC-biased quartered composite transform (QCT) eliminates filter-induced crosstalk and channelizes frequency selectivity into single-tap subchannels (Cinemre et al., 13 Apr 2025).

3. Implementation Architectures and Practical Systems

Systems implementing color-encoded illumination range from compact imaging setups to high-speed synchronized multi-camera arrays and large-scale LED installations.

Domain	Encoding Principle	Notable Implementation
Camera color constancy (Kim et al., 10 Apr 2025)	Planckian locus + CCM	CNN-guided U-Net using 8-D CFE
Volumetric imaging (Novikov et al., 29 Apr 2026)	Temporal color strobing	LED panels + dynamic Gaussian Splatting
SPD optimization (Yamaguchi et al., 2024)	LED basis, metamerism	15-LED stage fixtures, QP/SQP
Ghost imaging (Huang et al., 2017)	Masked R/G/B coding	Projector, PMT detector, TV-regularization
VLC (Gao et al., 2015, Cinemre et al., 13 Apr 2025)	RGB+DC+freq. constellations	Multi-channel bias and OFDM driver

All approaches require precise calibration of illumination (color space, intensity, timing) and detector response to enable accurate encoding and decoding. In dynamic scene settings, synchronization across devices is achieved via hardware triggers (e.g., Arduino-synchronized PWM strobes for multi-camera high-speed reconstruction (Novikov et al., 29 Apr 2026)).

4. Computational Decoding, Estimation, and Learning

Decoding color-encoded illumination typically involves direct inverses, convex optimization, or deep neural learns representations:

• Color constancy architectures: The CFE from the color-encoded locus steers a U-Net hypernetwork that dynamically generates convolutional filters and biases for histogram-based illumination estimation; loss is mean angular error in illumination direction (Kim et al., 10 Apr 2025).
• Compressive recovery: Multispectral ghost imaging reconstructs a fused measurement vector via TV-minimization and demultiplexes it into R/G/B by known binary masks, followed by further TV-regularized channel recovery (Huang et al., 2017).
• Spectral optimization: SPD mixing coefficients are found by quadratic or sequential quadratic programming, balancing colorimetric objectives with CRI/white-point constraints (Yamaguchi et al., 2024).
• Dynamic scene inversion: Volumetric scene trajectories are estimated by fitting time-dependent 3D Gaussian basis functions to match observed color mixtures under the color sequence, regularizing for smoothness (Novikov et al., 29 Apr 2026).
• Learned color normalization: Retinex-inspired two-branch networks (e.g., RLN²) explicitly disentangle chromaticity and luminance via cross-domain attention and multi-scale wavelet features, guided by HSV cues, with losses combining $\mathbf{R}$1 error and chromaticity consistency (Vasluianu et al., 4 Aug 2025).

5. Applications, Advantages, and Empirical Performance

Color-encoded illumination underpins applications in imaging, vision, communications, visualization, and display.

• Camera cross-domain color constancy: Achieves mean angular errors of 1.68°, 2.23°, and 2.32° on Cube⁺, Gehler-Shi, and NUS-8, respectively, outperforming prior models with compact architecture (Kim et al., 10 Apr 2025).
• High-speed volumetric capture: Realizes effective 600 FPS from 60 FPS camera hardware, reconstructs temporally sharp multi-view novel-views. MAE remains low for up to 20 subframes/sequence; accuracy degrades as color separation diminishes (Novikov et al., 29 Apr 2026).
• Visible light communication: QCT achieves SNR improvements of 26 dB, mean BER around $\mathbf{R}$2 at 10 dB SNR, and >2× increase in average illuminance relative to standard CSK, with maintained CRI and CCT (Cinemre et al., 13 Apr 2025).
• Metameric display control: Enables material-specific color metamorphosis (e.g., isochromatic textile effects) under white-appearing light, with constraints on observer-invariant color and white-point shifts (Yamaguchi et al., 2024).
• Compressive multispectral sensing: Recover color images with RMSE ≈ 3% using only 4000–6000 measurements for 81x81 images, security via coding masks (Huang et al., 2017).
• Dynamic event camera color recovery: DVS + flicker methods robustly reconstruct color imagery (RMSE ~45–47, PSNR ≈ 30–35 dB, SSIM ≈ 0.9) under a variety of lighting and scene conditions (Cohen et al., 2022).

6. Limitations, Tradeoffs, and Prospects

Several challenges and tradeoffs are inherent:

• Channel separability: Fidelity of encoding and decoding relies on sufficient “separation” in color space; as more time bins/colors are added, clusters in RGB space reduce invertibility and SNR (Novikov et al., 29 Apr 2026).
• Material and illumination invariance: Uniform-albedo and non-textured object assumptions can limit effectiveness; spatially varying or unknown reflectance spectra degrade performance.
• Color rendering and viewer constraints: For SPD optimization, the finite gamut of LED channels and observer perceptions of metamerism constrain the range of achievable color manipulations (Yamaguchi et al., 2024).
• Security and data reduction: Mask-based encoding in ghost imaging provides intrinsic security, but at the cost of increased computational complexity for demultiplexing (Huang et al., 2017).
• Computational requirements: High-dimensional optimization and learned models may require significant training or hardware investments.
• Visualization: Color-encoded shading for visualization is subject to potential hue/saturation distortion; advanced luminance-only approaches in LAB space preserve chromaticity for readability and data analysis (Chen et al., 23 Jul 2025).

Future directions include expanding LED/SPD bases for finer-grained control, integrating advanced learned priors in dynamic scene coding, improving calibration for textured/multi-albedo scenes, and extending to broader spectral regions (e.g., near-IR/UV), as well as deploying real-time adaptive control in architectural and entertainment lighting contexts.