100 Mfps ghost imaging with wavelength division multiplexing

Ghost imaging (GI) and single-pixel imaging (SPI) techniques enable image reconstruction without spatially resolved detectors, offering unique access to wide spectral ranges and challenging imaging environments. Yet, their adoption has been limited by the slow generation of mask patterns, which constrains achievable frame rates. Here, we demonstrate ultrafast GI that achieves a spatial-temporal information flux of 78.4 gigapixels per second across five wavelengths, which is at least two orders of magnitude larger than that reported for previous training-data-free GI approaches. This breakthrough is enabled by 25 GHz speckle pattern switching and allows parallelizing the pattern illumination using a wavelength-division multiplexing (WDM) technique. We show that the proposed approach is capable of reconstructing 28$\times$28-pixel images at the exposure time of 10 ns, achieving 100 megaframes per second (Mfps), and demonstrate the GI of a microsecond-scale dynamic event. This approach opens avenues for studying rapid processes in physics, chemistry, and biology, where conventional cameras are limited by detector bandwidth, readout speed, or cost.

💡 Research Summary

**
The paper presents a breakthrough in single‑pixel and ghost imaging (SPI/GI) by overcoming the long‑standing bottleneck of mask‑pattern generation speed. The authors achieve ultrafast ghost imaging through two complementary innovations: (1) a 25 GHz random speckle‑pattern switching scheme based on a high‑speed arbitrary waveform generator and a phase modulator that drives a 20‑meter multimode fiber, and (2) wavelength‑division multiplexing (WDM) that simultaneously projects five independent speckle sequences using five closely spaced lasers around 1550 nm. By combining these, the effective pattern‑switching rate reaches 125 GHz (25 GHz × 5), enabling a spatial‑temporal information flux (STIF) of 78.4 gigapixels per second for 28 × 28‑pixel images.

The system architecture consists of a multiplexer that combines the five laser sources, a phase modulator driven by a 25 GSa/s waveform to imprint pseudorandom phase codes, and a multimode fiber that converts the phase‑modulated light into rapidly varying speckle patterns. The speckle patterns for each wavelength are projected onto the target simultaneously. Reflected light is collected, demultiplexed, and detected by an array of fast photodetectors, producing time‑domain signals yλ(t) for each wavelength.

To reconstruct images, the authors first estimate the measurement matrix S by displaying 2 500 random patterns on a DMD and recording the corresponding time‑domain responses for each wavelength. The matrix S has dimensions (K M) × N, where K = 5 wavelengths, M is the number of time samples per exposure, and N = 28 × 28 = 784 pixels. A naïve pseudo‑inverse reconstruction (x̂ = S†y) suffers from low rank and noise, especially at very short exposure times. Therefore, they introduce a self‑supervised deep learning framework called Ghost Imaging using Deep neural‑network Constraint (GIDC). GIDC takes the pseudo‑inverse estimate as input, refines it with a U‑Net‑style convolutional neural network, and enforces a total‑variation regularizer in the loss function L = ‖ŷ − y‖² + TV(x̂). This approach yields high‑fidelity reconstructions without any external training dataset, making it suitable for previously unseen dynamic events.

Experimental validation uses MNIST handwritten digits (28 × 28) displayed on a DMD. With a single wavelength (K = 1), reconstruction quality (SSIM) degrades sharply for exposure times below 10 ns. However, increasing the number of wavelengths to five dramatically improves SSIM, achieving values above 0.8 even at 1 ns exposure. This demonstrates that parallel, uncorrelated speckle patterns from multiple wavelengths effectively increase pattern diversity and compensate for the reduced photon budget at ultrashort exposures.

A proof‑of‑concept video captures a microsecond‑scale switching event: a 16 × 16 diamond pattern on the DMD is switched to a heart shape. The continuous time‑domain signal is segmented into 10 ns intervals, and each segment is reconstructed using the five‑wavelength WDM‑GI pipeline. The resulting video runs at 100 Mfps, clearly showing the transition with an average SSIM of ~0.85. Small residual fluctuations after the transition are attributed to mechanical ringing of the DMD mirror.

A comparative table places the proposed method against prior high‑speed SPI/GI techniques, including mechanical (rotating disks), electronic (LED, DMD), and optical (dual‑comb) approaches. The new system outperforms previous work by two to three orders of magnitude in STIF and achieves a frame rate of 100 Mfps, while requiring no pre‑collected training data for reconstruction.

The authors acknowledge current limitations: the demonstrated resolution is limited to 28 × 28 pixels, the setup relies on expensive high‑speed waveform generators and multiple narrow‑linewidth lasers, and the measurement matrix must be calibrated beforehand. Future work could explore scaling to larger detector arrays, integrating on‑chip photonic modulators for more compact speckle generation, and developing adaptive calibration schemes to reduce overhead.

In conclusion, the combination of 25 GHz random speckle switching and wavelength‑division multiplexing enables ultrafast, training‑data‑free ghost imaging capable of recording microsecond‑scale dynamics at 100 Mfps. This opens new possibilities for studying rapid physical, chemical, and biological processes that are inaccessible to conventional cameras limited by detector bandwidth, readout speed, or cost.