Single Pixel Image Classification using an Ultrafast Digital Light Projector

Aisha Kanwal; Graeme E. Johnstone; Fahimeh Dehkhoda; Johannes H. Herrnsdorf; Robert K. Henderson; Martin D. Dawson; Xavier Porte; Michael J. Strain

cs.CVMar 12, 2026

Single Pixel Image Classification using an Ultrafast Digital Light Projector

Authors: Aisha Kanwal, Graeme E. Johnstone, Fahimeh Dehkhoda, Johannes H. Herrnsdorf, Robert K. Henderson, Martin D. Dawson, Xavier Porte, Michael J. Strain

Organizations: Institute of Photonics, University of Strathclyde, 99 George Street, Glasgow, G1 1RD, UK · School of Engineering, University of Edinburgh, Sanderson Building, Robert Stevenson Road, Edinburgh, EH9 3FB, UK

Abstract

Pattern recognition and image classification are essential tasks in machine vision. Autonomous vehicles, for example, require being able to collect the complex information contained in a changing environment and classify it in real time. Here, we experimentally demonstrate image classification at multi-kHz frame rates combining the technique of single pixel imaging (SPI) with a low complexity machine learning model. The use of a microLED-on-CMOS digital light projector for SPI enables ultrafast pattern generation for sub-ms image encoding. We investigate the classification accuracy of our experimental system against the broadly accepted benchmarking task of the MNIST digits classification. We compare the classification performance of two machine learning models: An extreme learning machine (ELM) and a backpropagation trained deep neural network. The complexity of both models is kept low so the overhead added to the inference time is comparable to the image generation time. Crucially, our single pixel image classification approach is based on a spatiotemporal transformation of the information, entirely bypassing the need for image reconstruction. By exploring the performance of our SPI based ELM as binary classifier we demonstrate its potential for efficient anomaly detection in ultrafast imaging scenarios.

Figures & tables

Figure 1: Schematic diagram of our single pixel image classification (SPIC) framework. Each Hadamard pattern in the imaging base is projected onto the object. The superposition of both images is converted into a scalar photocurrent and recorded by a real time oscilloscope. The sequence of all patterns is processed by an artificial neural network whose output is already the object’s class.

Figure 2: Experimental setup for single pixel image classification. The microLED projector generates the sequence of Hadamard patterns that illuminate the image presented on the Digital Micromirror Device (DMD). The Single Pixel detector collects the superposition of each image pair and transforms it into a scalar photocurrent registered by the real time oscilloscope. The full time series corresponding to each image is post-processed by a software encoded ML model.

Figure 3: Single pixel measurement and reconstruction process. (a) Original and binarized MNIST digit used as the target object for structured illumination. (b) Illustration of the SPI process: The upper panels show an optically encoded binarized MNIST digit and its corresponding photodetected time series. The panels below show a magnification of this time series, superposing the corresponding Hadamard patterns and the reconstructed MNIST digit.

Figure 4: Basic architectures and training of the two machine learning models considered in this work. Diagrams (a) and (b) depict the extreme learning machine model used for multiclass and one-vs-all binary classification with fixed random input weights and output training only using ridge regression. Diagrams (c) and (d) depict a feed-forward deep neural network for multiclass classification with trainable weights using gradient descent.

Figure 5: (a) Classification performance of the multiclass ELM model versus hidden neurons at $\alpha=1$ . (b) Normalized confusion matrix (%) for the multiclass ELM model using $1000$ hidden neurons at $\alpha=1$ .

Figure 6: (a) Classification performance of the one-vs-all ELM binary classifier at

\alpha=1

with

1000

hidden neurons. (b) Receiver operating characteristics (ROC) curves for each class in binary ELM classifier. The different values for the area under the curve (AUC) are all close to 1, indicating strong separability for all classes.

Figure 7: Validation accuracy for different subsets of Had12 after model training for 300 epochs. The bandwidths at which the images are classified are depicted on the top axis.

Figure 8: Validation accuracy learning curves over 300 epochs for a DNN on MNIST variants and Had12 subsets. (a) Numerical simulation of the original (grayscale) MNIST and binarized MNIST, and experimental MNIST encoded with full Had12 set. (b) Binary MNIST encoded with the full Had12 and with the first

\tfrac{1}{2}

\tfrac{1}{4}

\tfrac{1}{8}

, and

\tfrac{1}{16}

patterns of the Had12 set. (c) Binary MNIST encoded with the full Had12 and with the last

\tfrac{1}{2}

\tfrac{1}{4}

\tfrac{1}{8}

, and

\tfrac{1}{16}

patterns of the Had12 set. (d) Binary MNIST encoded with the full Had12 set and with randomly selected

\tfrac{1}{2}

\tfrac{1}{4}

\tfrac{1}{8}

, and

\tfrac{1}{16}

pairs of the Had12 set.

Figure 9: (a) Examples of MNIST digits injected with additive uncorrelated Gaussian noise at various noise intensities $\sigma=0.1,0.5,1.0$ . (b) Numerical validation accuracy of a feedforward DNN on binary MNIST dataset. Additive uncorrelated Gaussian noise is injected in the input layer, where $\sigma=\sqrt{2D}$ shows the noise intensity.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

P_{i,j}=H_{N}(:,i)\,H_{M}(j,:)

Appendix

Algorithm 1 Generation of Hadamard Pattern Set

Figure 10: Examples of Hadamard patterns with binary intensity values encoded in white (+1) and black (-1). Panels (a), (b) and (c) correspond to patterns in category 1 showing one-dimensional spatial distribution only (cf. explanation in the main text). Panels (d), (e) and (f) illustrate patterns in category 2, i.e. showing two-dimensional spatial distribution. (g) Validation accuracy of the experimental binarized MNIST using only Hadamard12 categories over 300 training epochs.

Explore similar work

Compressive single-pixel imaging via a wavelength-multiplexed spatially incoherent diffractive optical processor
Mar 23, 2026Xiao Wang, Yiyang Wu, Yuntian Wang +8Computational Imaging Photonic Computing
Electrical and Computer Engineering Department, University of California, Los Angeles, CA, 90095, USA · Bioengineering Department, University of California, Los Angeles, CA, 90095, USA · California NanoSystems Institute (CNSI), University of California, Los Angeles, CA, 90095, USA
Machine vision with small numbers of detected photons per inference
Mar 25, 2026Shi-Yuan Ma, Jérémie Laydevant, Mandar M. Sohoni +3Image Classification Hardware-Software Co-Design
School of Applied and Engineering Physics, Cornell University, Ithaca, NY 14853, USA · Present address: Research Laboratory of Electronics, Massachusetts Institute of Technology, Cambridge, MA 02139, USA. · NTT Physics and Informatics Laboratories, NTT Research, Inc., Sunnyvale, CA 94085, USA +2
LUMIN: Lightweight Universal Manufacturing Inspection Network for Anomaly Detection
Sep 4, 2026Pengfei YangIndustrial Anomaly Detection Anomaly Detection
Intelligent Precision Instrument, Shenzhen, Guangdong, China

Mar 23, 2026J/K move · Enter open · S save

Compressive single-pixel imaging via a wavelength-multiplexed spatially incoherent diffractive optical processor

Xiao Wang, Yiyang Wu, Yuntian Wang, Md Sadman Sakib Rahman+7

Electrical and Computer Engineering Department, University of California, Los Angeles, CA, 90095, USA · Bioengineering Department, University of California, Los Angeles, CA, 90095, USA · California NanoSystems Institute (CNSI), University of California, Los Angeles, CA, 90095, USA

Computational Imaging Photonic Computing

Open paper AskPDF arXiv

Despite offering high sensitivity, a high signal-to-noise ratio, and a broad spectral range, single-pixel imaging (SPI) is limited by low measurement efficiency and long data-acquisition times. To address this, we propose a wavelength-multiplexed, spatially incoherent diffractive optical processor combined with a compact/shallow digital artificial neural network (ANN) to implement compressive SPI. Specifically, we model the bucket detection process in conventional SPI as a linear intensity transformation with spatially and spectrally varying point-spread functions. This transformation matrix is treated as a learnable parameter and jointly optimized with a shallow digital ANN composed of 2 hidden nonlinear layers. The wavelength-multiplexed diffractive processor is then configured via data-free optimization to approximate this pre-trained transformation matrix; after this optimization, the diffractive processor remains static/fixed. Upon multi-wavelength illumination and diffractive modulation, the target spatial information of the input object is spectrally encoded. A single-pixel detector captures the output spectral power at each illumination band, which is then rapidly decoded by the jointly trained digital ANN to reconstruct the input image. In addition to our numerical analyses demonstrating the feasibility of this approach, we experimentally validated its proof-of-concept using an array of light-emitting diodes (LEDs). Overall, this work demonstrates a computational imaging framework for compressive SPI that can be useful in applications such as biomedical imaging, autonomous devices, and remote sensing.

Abstract

Explore similar work

Compressive single-pixel imaging via a wavelength-multiplexed spatially incoherent diffractive optical processor

Machine vision with small numbers of detected photons per inference

LUMIN: Lightweight Universal Manufacturing Inspection Network for Anomaly Detection