Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Scalable dataset acquisition for data-driven lensless imaging

Clara S. Hung, Leyla A. Kabuli, Vasilisa Ponomarenko, Laura Waller

TL;DR

The paper tackles the data bottleneck in data-driven lensless imaging by introducing a parallel acquisition pipeline that captures measurements from two lensless imagers and a ground-truth lensed camera under identical conditions, with computational ground-truth registration. It delivers an open-access 25,000-image dataset, accompanied by reproducible hardware and open-source synchronization software, PSF calibration, and alignment techniques. Ground-truth reconstructions leverage $200$ iterations of FISTA, and PSF autocorrelation analyses indicate improved resolution and reduced sidelobes compared with prior work. This dataset and framework enable robust training of reconstruction networks and end-to-end system design for lensless imaging, including the potential use of transformer-based approaches for improved performance.

Abstract

Data-driven developments in lensless imaging, such as machine learning-based reconstruction algorithms, require large datasets. In this work, we introduce a data acquisition pipeline that can capture from multiple lensless imaging systems in parallel, under the same imaging conditions, and paired with computational ground truth registration. We provide an open-access 25,000 image dataset with two lensless imagers, a reproducible hardware setup, and open-source camera synchronization code. Experimental datasets from our system can enable data-driven developments in lensless imaging, such as machine learning-based reconstruction algorithms and end-to-end system design.

Scalable dataset acquisition for data-driven lensless imaging

TL;DR

The paper tackles the data bottleneck in data-driven lensless imaging by introducing a parallel acquisition pipeline that captures measurements from two lensless imagers and a ground-truth lensed camera under identical conditions, with computational ground-truth registration. It delivers an open-access 25,000-image dataset, accompanied by reproducible hardware and open-source synchronization software, PSF calibration, and alignment techniques. Ground-truth reconstructions leverage iterations of FISTA, and PSF autocorrelation analyses indicate improved resolution and reduced sidelobes compared with prior work. This dataset and framework enable robust training of reconstruction networks and end-to-end system design for lensless imaging, including the potential use of transformer-based approaches for improved performance.

Abstract

Data-driven developments in lensless imaging, such as machine learning-based reconstruction algorithms, require large datasets. In this work, we introduce a data acquisition pipeline that can capture from multiple lensless imaging systems in parallel, under the same imaging conditions, and paired with computational ground truth registration. We provide an open-access 25,000 image dataset with two lensless imagers, a reproducible hardware setup, and open-source camera synchronization code. Experimental datasets from our system can enable data-driven developments in lensless imaging, such as machine learning-based reconstruction algorithms and end-to-end system design.
Paper Structure (7 sections, 3 figures)

This paper contains 7 sections, 3 figures.

Table of Contents

  1. INTRODUCTION
  2. Methods
  3. Hardware
  4. Software
  5. System Calibration & Analysis
  6. Discussion
  7. Conclusion

Figures (3)

  • Figure 1: Parallel dataset acquisition setup:(a) A diagram of the hardware system. Images are displayed in parallel with two lensless imagers capturing lensless measurements and a lensed camera, which captures ground truth. (b) Experimental image of the setup, with two images displayed in parallel for simultaneous capture by all imagers.
  • Figure 2: Software pipeline for image and camera control. The displayed images pass through each imaging system to generate measurements. Lensless measurements are computationally reconstructed while the ground truth lensed measurement is undistorted. Reconstructed images are computationally aligned to the undistorted ground truth image.
  • Figure 3: System calibration PSFs. (a) Our DiffuserCam point spread function (PSF). (b) Our Random Multi-focal Lenslet (RML) PSF. (c) An autocorrelation comparison between our DiffuserCam PSF, our RML PSF, and previous work that used a Gaussian diffuser Monhakova et al. Kristina. Both our DiffuserCam PSF (blue) and RML PSF (purple) have a sharper main lobe and lower sidelobes, indicating higher resolution and higher signal-to-noise ratio (SNR) compared to Monhakova et al. Kristina. (d) Sample reconstructions of DiffuserCam and RML measurements and undistorted ground truth images acquired using our system.