Scalable dataset acquisition for data-driven lensless imaging

Technical Summary: Scalable Dataset Acquisition for Data-Driven Lensless Imaging

Problem Statement
The advancement of lensless imaging, particularly through machine learning-based reconstruction algorithms and end-to-end system design, is currently hindered by a scarcity of large-scale, high-quality experimental datasets. While simulated data exists, it fails to capture the non-idealities of real-world imaging systems. Existing experimental datasets face significant limitations: some are too small (e.g., 10,000 images) to train state-of-the-art algorithms, others are not publicly available, and those that are large enough (e.g., 25,000 images) often contain data from only a single optical configuration (single phase mask). Furthermore, acquiring such data is challenging due to the need for precise alignment, calibration, and the lack of flexible, automated pipelines capable of capturing measurements from multiple systems in parallel under identical conditions.

Methodology
The authors propose a reproducible, open-source data acquisition pipeline comprising a modular hardware system and a synchronized software framework.

• Hardware Setup: The system utilizes a portable monitor (INNOCN 13A1F) to display ground truth images from the MIRFLICKR-25000 dataset. These images are captured simultaneously by two distinct lensless imaging systems and a reference lensed camera.
  • Lensless Systems: Two different configurations are tested side-by-side: a DiffuserCam (using a Gaussian diffuser, 4.5 mm from the sensor) and a Random Multi-focal Lenslet (RML) system (using a phase mask, 18 mm from the sensor).
  • Ground Truth: A lensed camera (Basler daA1920-160uc with a 6 mm lens) captures the undistorted scene. To ensure identical optical paths, the lensed camera is aligned via a 50/50 beamsplitter with the DiffuserCam.
  • Environment: The setup is enclosed with opaque black dividers to control stray light, and the imagers are mounted on a custom 3D-printed frame with a 135 mm separation.
• Software Framework: A custom Python-based pipeline automates the process.
  • Synchronization: The monitor displays images sequentially, triggering all cameras synchronously with specific delays (200 ms between cameras, 500 ms between images) to ensure continuous acquisition.
  • Calibration: Exposure levels are heuristically calibrated to prevent saturation. Lens distortion in the ground truth camera is corrected using OpenCV calibration grids.
  • Alignment: To achieve pixel-to-pixel correspondence, a homography is learned using the Kornia package to computationally align the reconstructed lensless images with the undistorted ground truth, correcting for minor shifts unadjustable by optical alignment.
  • Reconstruction: During calibration, images are reconstructed using 200 iterations of the FISTA algorithm to visualize system performance.

Key Contributions

1. Open-Access Dataset: The authors release a dataset of 25,000 images containing paired measurements from two different lensless imaging systems (DiffuserCam and RML) and their corresponding ground truth images.
2. Reproducible Pipeline: The paper provides the hardware schematics (including 3D-printed mount designs) and open-source software code (available via GitHub) to replicate the parallel data acquisition setup.
3. Parallel Acquisition Capability: The system demonstrates the ability to capture data from multiple lensless systems under identical imaging conditions simultaneously, a capability previously lacking in public datasets.

Results and System Analysis
The authors characterize the system's performance by analyzing the Point Spread Functions (PSFs) of both the DiffuserCam and RML configurations.

• Resolution and SNR: Autocorrelation analysis of the captured PSFs reveals that both systems in this setup exhibit narrower main lobes and lower sidelobes compared to a previous state-of-the-art DiffuserCam dataset (Monhakova et al.). This indicates superior resolution and signal-to-noise ratio (SNR).
• System Comparison: The RML configuration demonstrates minimal sidelobes compared to the DiffuserCam, suggesting improved image quality and SNR for that specific phase mask.
• Reconstruction Quality: Sample reconstructions using FISTA confirm that the system captures high-quality measurements suitable for training advanced algorithms.

Significance and Claims
The paper claims that this work directly addresses the data bottleneck in lensless imaging research. By providing a large-scale, multi-system, experimental dataset with paired ground truth, the authors enable:

• The training of data-driven reconstruction algorithms (e.g., deep learning models like transformers) that require tens of thousands of images.
• Fair, quantitative comparisons between different lensless imaging systems and phase masks.
• Information-theoretic analyses and end-to-end system design optimizations that rely on real-world experimental data rather than simulations.

The authors maintain a modest scope, noting that while the current demonstration uses specific phase masks, the modular nature of the hardware allows for future extension to other configurations, 3D datasets, and video acquisition. The primary contribution is the facilitation of further data-driven developments in the field through accessible, high-quality experimental resources.