Skip to content

Single Pixel Image Classification using an Ultrafast Digital Light Projector

Conference: ICLR 2026 arXiv: 2603.12036 Code: None Area: Autonomous Driving Keywords: Single-pixel imaging, image classification, microLED, Hadamard patterns, extreme learning machine

TL;DR

This paper presents an experimental single-pixel imaging (SPI) system based on a microLED-on-CMOS ultrafast digital light projector, combined with low-complexity machine learning models (ELM and DNN) to achieve sub-millisecond image encoding and kHz-rate image classification. The system attains >90% accuracy on the MNIST dataset and >99% AUC in binary classification scenarios.

Background & Motivation

  1. Background: Machine vision is a mature technology embedded in autonomous agents such as self-driving vehicles; however, the operational bandwidth of conventional digital cameras is becoming a bottleneck. Event cameras reduce data volume in dynamic scenes but are constrained to the visible and near-infrared spectrum.
  2. Limitations of Prior Work:
    • Conventional SPI systems employ DMDs (digital micromirror devices) for pattern generation, which are limited by mechanical switching speeds (~\(10^4\) fps), keeping overall imaging rates comparable to standard CMOS cameras (\(\lesssim 10^2\) Hz).
    • Most existing single-pixel image classification (SPIC) work relies on simulation or low-speed experiments, lacking true ultrafast optical experimental validation.
    • The image reconstruction step introduces additional latency and computational overhead.
  3. Key Challenge: SPI requires projecting long sequences of patterns to acquire sufficient information, and projection speed constitutes the bandwidth bottleneck; compressed sensing can reduce the number of patterns but at the cost of classification accuracy.
  4. Goal: To experimentally validate a single-pixel image classification system based on ultrafast microLED projection that performs direct classification on photodetector time series without image reconstruction.
  5. Key Insight: Leveraging the ~100× faster switching speed of microLED arrays compared to DMDs to bypass image reconstruction entirely and classify directly on spatiotemporally transformed data.
  6. Core Idea: Reformulating image classification from the spatial domain to the spatiotemporal domain—each image is encoded as a light-intensity time series and classified directly by a low-complexity ML model.

Method

Overall Architecture

Hadamard pattern sequence → high-speed microLED projector → target image displayed on DMD → single-pixel detector captures superimposed light-intensity signal → real-time oscilloscope records time series → ML model performs direct classification (no reconstruction required).

Key Designs

  1. Ultrafast Single-Pixel Imaging System:

    • Core hardware: 128×128 microLED-on-CMOS array, 30×30 μm² pixels, 50 μm pitch, supporting MHz-rate frame refresh.
    • Projects a 12×12 Hadamard pattern set (Had12) at 330,000 fps in global shutter mode.
    • Image reconstruction formula: \(I_{(x,y),M} = \frac{1}{M}\sum_{m=1}^{M} S_m P_{(x,y),m}\), where \(S_m\) is the differential signal between each pair of complementary Hadamard patterns.
    • Binarized MNIST images are displayed on a DMD (1024×768 resolution).
    • An Onsemi SiPM single-pixel detector captures light intensity, recorded by a 1 GHz bandwidth oscilloscope.
    • Design Motivation: The MHz-level switching speed of microLEDs overcomes the mechanical limitations of DMDs, enabling true kHz-rate SPI.

  2. Extreme Learning Machine (ELM) Classifier:

    • Single hidden-layer neural network with randomly initialized and fixed input weights.
    • Hidden layer output: \(H = f(XW_{\text{in}} + b)\), using ReLU activation.
    • Output weights solved in closed form via Ridge regression: \(\beta = (H^\top H + \alpha I)^{-1} H^\top T\).
    • Multi-class prediction uses \(\hat{y} = \max(Y)\); binary classification uses a threshold of 0.5.
    • Regularization parameter \(\alpha = 1.0\).
    • Design Motivation: ELM training is extremely fast (no iterative optimization); inference requires only 31 μs/image, making it well-suited for ultrafast scenarios.

  3. Deep Neural Network (DNN) Classifier:

    • Feedforward DNN: input layer (286-dimensional) → three hidden layers with decreasing width + ReLU → softmax output.
    • Adam optimizer, sparse categorical cross-entropy loss, trained for 300 epochs.
    • Inference time: 73 μs/image.
    • Design Motivation: Serves as a higher-complexity baseline for comparison with ELM, exploring the accuracy–speed trade-off.

  4. Hadamard Pattern Subset Optimization:

    • Low-index (low spatial frequency) Hadamard patterns are found to carry more classification-relevant information.
    • Using only the first 1/4 of patterns preserves ≃78% classification accuracy.
    • Cat1 (first 44 patterns) varies along a single spatial axis and captures coarse features; Cat2 (patterns 45–288) varies along two spatial directions and captures finer features.
    • Design Motivation: Reducing the number of projected patterns proportionally increases the effective imaging bandwidth.

Loss & Training

  • ELM: closed-form Ridge regression solution, \(\alpha = 1.0\), no iterative training.
  • DNN: Adam optimizer, sparse categorical cross-entropy loss, 300 epochs.
  • Data: MNIST dataset (60K training, 10K testing), binarized and rescaled to fill the full DMD surface.

Key Experimental Results

Main Results

Method / Configuration Accuracy (%) Inference Speed Notes
DNN + full Had12 (experimental) >90 73 μs/image 1.2 kfps frame rate
ELM + full Had12 (experimental) 87.37 31 μs/image 2× faster than DNN
DNN + binarized MNIST (simulation) 97.50 — Theoretical upper bound
ELM + binarized MNIST (simulation) 93.32 — ELM upper bound
ELM binary classification (one-vs-all) AUC >99% — Anomaly detection

Ablation Study

Configuration Accuracy (%) Notes
Full Had12 (DNN) >90 All 144 patterns
First 1/2 Had12 ~86 Minor accuracy drop
First 1/4 Had12 ~78 Acceptable accuracy, ×4 bandwidth
First 1/8 Had12 ~68 Significant accuracy drop
Last 1/2 Had12 ~75 High-frequency patterns less informative
Random 1/2 Had12 ~82 Between first/last halves
Gaussian noise σ=0.1 >95 Minimal noise impact
Gaussian noise σ=0.5 >95 Convergence maintained
Gaussian noise σ=1.0 ~85 Notable drop and variance

Key Findings

  • The primary cause of accuracy degradation is not reduced effective SNR but spatial information loss due to compressed sensing.
  • Low spatial frequency Hadamard patterns contribute most to classification; high-frequency patterns are less informative.
  • Although ELM achieves lower accuracy than DNN, its inference speed is 2× faster, making it suitable for extreme real-time scenarios.
  • When the number of patterns is reduced, the DNN exhibits longer vanishing gradient phases, consistent with the nature of compressed inputs.
  • In binary classification, AUC approaches 1.0, indicating suitability for anomaly detection in rapidly changing scenes.

Highlights & Insights

  • This work provides the first experimental validation of single-pixel image classification at kHz frame rates, surpassing the speed limitations of conventional imaging systems.
  • Bypassing image reconstruction entirely greatly simplifies the system pipeline and reduces latency.
  • The minimalist design of the ELM model aligns well with the demands of ultrafast scenarios, offering fast training, fast inference, and low overhead.
  • The frequency-domain analysis of Hadamard pattern subsets offers practical guidance for compression strategy design.
  • The comparative experiments on noise versus compressed sensing provide valuable theoretical insights.

Limitations & Future Work

  • Validation is limited to the MNIST dataset, which is relatively simple and far from real-world machine vision scenarios.
  • The 12×12 Hadamard pattern resolution is low, limiting the system's ability to resolve complex images.
  • The storage depth of the current FPGA board constrains the size of the pattern set.
  • The SiPM detector and oscilloscope on the sensing side are difficult to miniaturize and integrate.
  • More complex ML models (e.g., CNNs) and larger-scale datasets have not been explored.
  • Substantial work remains to transfer the approach from MNIST to practical autonomous driving scenarios.
  • Compressed sensing theory provides the mathematical foundation for reducing the number of projected patterns.
  • The application of microLED arrays in analog optical computing underscores their central role in next-generation optical computing architectures.
  • Reconstruction-free SPIC methods have advanced rapidly in recent years; this paper represents the fastest experimentally validated instance among them.
  • The combination of low-complexity models such as ELM and reservoir computing with optical hardware is a promising research direction.
  • Single-pixel imaging technology holds unique advantages in non-visible spectral bands (terahertz, ultraviolet).

Rating

  • Novelty: ⭐⭐⭐⭐
  • Experimental Thoroughness: ⭐⭐⭐
  • Writing Quality: ⭐⭐⭐⭐
  • Value: ⭐⭐⭐