Generative Adversarial Network for Superresolution Imaging through a Fiber

Open Access

Generative Adversarial Network for Superresolution Imaging through a Fiber

Wei Li, Ksenia Abrashitova, Gerwin Osnabrugge, and Lyubov V. Amitonova

Phys. Rev. Applied 18, 034075 – Published 27 September 2022

Abstract

A multimode fiber represents the ultimate limit in miniaturization of imaging endoscopes. However, such a miniaturization usually comes as a cost of a low spatial resolution and a long acquisition time. Here we propose a fast superresolution-fiber-imaging technique employing compressive sensing through a multimode fiber with a data-driven machine-learning framework. We implement a generative adversarial network (GAN) to explore the sparsity inherent to the model and provide compressive reconstruction images that are not sparse in a representation basis. The proposed method outperforms other widespread compressive imaging algorithms in terms of both image quality and noise robustness. We experimentally demonstrate machine-learning ghost imaging below the diffraction limit at a sub-Nyquist speed through a thin multimode fiber probe. We believe that this work has great potential in applications in various fields ranging from biomedical imaging to remote sensing.

Received 15 March 2022
Revised 17 August 2022
Accepted 26 August 2022

DOI:https://doi.org/10.1103/PhysRevApplied.18.034075

Published by the American Physical Society under the terms of the Creative Commons Attribution 4.0 International license. Further distribution of this work must maintain attribution to the author(s) and the published article's title, journal citation, and DOI.

Published by the American Physical Society

Physics Subject Headings (PhySH)

Imaging & optical processing Optics & lasers Ultrafast optics

Artificial neural networks Optical fibers

Ghost imaging Machine learning Super-resolution techniques

Atomic, Molecular & OpticalCondensed Matter, Materials & Applied Physics

Authors & Affiliations

Wei Li ^1,*, Ksenia Abrashitova ¹, Gerwin Osnabrugge¹, and Lyubov V. Amitonova ^1,2

¹Advanced Research Center for Nanolithography (ARCNL), Science Park 106, Amsterdam 1098 XG, Netherlands
²LaserLaB, Department of Physics and Astronomy, Vrije Universiteit Amsterdam, De Boelelaan 1081, Amsterdam 1081 HV, Netherlands

^*w.li@arcnl.nl

Article Text

Click to Expand

References

Click to Expand

Issue

Vol. 18, Iss. 3 — September 2022

Subject Areas

Reuse & Permissions

Author publication services for translation and copyediting assistance advertisement

Images

Figure 1
(a) Simplified representation of machine-learning fiber-based ghost imaging. Different speckle patterns are sequentially projected on the sample. The corresponding total transmitted intensity for each speckle pattern is measured by a bucket detector. The GAN-based algorithm reconstructs the image from the intensity sequence measured by the bucket detector and speckle patterns measured by the camera. (b) Examples of simulated speckle patterns with different cutoff frequencies, $ν$ . (c) Schematic representation of the experimental setup. BS, beam splitter. (d) Scanning grid. (e) Incoherent image of the output facet. (f) Coherent image of the output facet. The scale bar is $10 μ m$ . (g) Probability density function of correlation coefficients for the experimental set of speckle patterns.
Reuse & Permissions
Figure 2
Illustration of machine-learning fiber-based ghost imaging via GAN: (a) the training process and (b) the imaging step. SP, speckle patterns; BDS, bucket detector signal.
Reuse & Permissions
Figure 3
(a) Results of the simulations performed on a handwritten digit “3” (randomly selected from the testing dataset) and handwritten letters “ $e$ ” and “ $x$ ” with $ν = 0.2$ and $m = 40$ , 100 and 200, respectively. The first column is the ground truth. The second column is the diffraction-limited image. Images reconstructed by the BP algorithm and by the GAN under a different number of measurements are shown from third to eighth column. (b) Average power spectrum for the reconstructed images for the cutoff frequency $ν = 0.2$ , $m \geq 40$ . The violet dashed circle corresponds to $ν = 0.2$ , the white dashed circle marks the spatial frequency where the amplitude of the power spectrum decreases 4 orders of magnitude. All spatial frequencies are expressed in the units of half of the spatial sampling frequency.
Reuse & Permissions
Figure 4
Correlation between the reconstructed and ground truth images for diffraction-limited imaging, the BP (BPDN with noise) algorithm and GAN framework as functions of cutoff frequency $ν$ and the number of measurement, averaged over the randomly selected digit of each kind ( $0 - 9$ ) for noiseless case (a),(b),(f) and for different noise levels: 40 dB (c),(g), 35 dB (d),(h), and 30 dB (e),(i) . The gray scale bar shows the $0$ – $0.9$ range and the colorbar shows the detailed $0.9$ – $1$ region. The Nyquist limit is indicated by the red dashed line.
Reuse & Permissions
Figure 5
Example of one validation image with $m = 300$ . (a) is the bright-field image of the sample engraved in the aluminum film, which is randomly chosen from the testing dataset. (b) is the diffraction-limited imaging of (a) when $ν = 0.2$ . (c),(d),(e) show the reconstructed image from 300 measurements with BP, BPDN, and GAN, respectively. The average performance of reconstruction on the real samples of BP, BPDN, and GAN algorithms is shown in (f).
Reuse & Permissions

Physical Review Applied