100 Mfps ghost imaging with wavelength division multiplexing

TL;DR

This work tackles the temporal bottleneck in ghost imaging by introducing an optical system that combines 25 GHz speckle pattern switching with wavelength-division multiplexing across five channels and a self-supervised, training-data-free reconstruction. It delivers 28\times 28-pixel images at 100 Mfps, corresponding to a spatial-temporal information flux of $78.4$ Gpps across five wavelengths, and demonstrates microsecond-scale video of a dynamic event. The approach outperforms prior training-data-free GI methods in throughput and suggests a pathway toward scalable ultrafast computational imaging with broad impact for observing fast processes in physics, chemistry, and biology. By enabling high-throughput, label-free reconstruction without training data, this method broadens the operational regime of SPI/GI and sets a foundation for future multispectral, ultrafast imaging systems.

Abstract

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.

Figures (7)

• Figure 1: Conceptual schematic of the proposed wavelength-multiplexed ghost imaging (WDM-GI) system. (a) Schematic of the WDM-GI system. DEMUX: demultiplexer; PDs: photodetectors. (b) Random speckle pattern projector. MUX: multiplexer; PM: phase modulator; MMF: multimode fiber. (c) Reconstruction model. NN: neural network.
• Figure 2: Reconstruction model. The network first computes a simplified reconstructed image from the pseudo-inverse of the measurement matrix $\boldsymbol{S}$, then refines the image using a CNN to suppress noise and enhance structural fidelity. The U-Net architecture is trained such that the output $\hat{\boldsymbol{y}} = \boldsymbol{S}\hat{\boldsymbol{x}}$ matches the target $\boldsymbol{y}$.
• Figure 3: (a) Measured random speckle patterns for $\lambda_1~=~$1550.12, $\lambda_2~=~$1549.32, $\lambda_3~=~$1548.52, $\lambda_4~=~$1547.72, and $\lambda_5~=~$1546.92 nm. (b) Correlation matrix between the measured speckle patterns for a single wavelength (1546.92 nm). Both axes are expressed in time (ns). (c) Average correlation matrix across five wavelengths.
• Figure 4: Reconstruction results with varying exposure times ($K$ = 1). (a) Measured time-domain signal and (b) its reconstructed images. Exposure time $T$: 1.0 - 100.0 ns. $\lambda = 1546.92$ nm. (c) SSIM as a function of $T$.
• Figure 5: Reconstruction results based on the WDM-GI. Exposure time $T$: 1.0, 5.0, 10.0 ns, number of wavelengths $K$ = 1 - 5, (a) Reconstructed images, (b) SSIM versus $K$ for each $T$. The gain from increasing $K$ is larger at shorter $T$. Our WDM-GI enables high-quality reconfiguration with a short exposure time $T \approx 10$ ns, corresponding to 100 Mfps.