Scalable dataset acquisition for data-driven lensless imaging

Clara S. Hung, Leyla Kabuli, V. A. Ponomarenko, Laura WallerarXiv 2025 · DOI · arXiv

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.

Introduction

Lensless imaging offers a path toward ultra-compact cameras by replacing traditional, bulky lenses with thin phase masks—such as textured plastic diffusers—placed directly on the image sensor. This configuration produces a multiplexed, encoded measurement that requires computational reconstruction to recover a sharp image. While promising, lensless systems often struggle with image quality.

The field is increasingly turning toward data-driven reconstruction methods, specifically deep learning. However, these models are notoriously data-hungry, often requiring tens of thousands of training examples. Existing datasets are frequently limited in size, proprietary, or restricted to a single mask type, hindering comparative research. To address this, researchers at UC Berkeley have developed a scalable, open-source pipeline to capture massive experimental datasets simultaneously across multiple lensless sensors and a lensed "ground truth" camera.

Methodology and Hardware Design

The core contribution is a hardware and software framework designed for high-throughput, parallelized data acquisition.

Hardware Setup

The system utilizes a portable monitor to display source images. A 50/50 beamsplitter is employed to allow a traditional lensed camera (providing ground truth) and a lensless imager to observe the exact same scene. To increase throughput and ensure fair comparison, multiple lensless imagers are mounted side-by-side.

The parallelized data acquisition setup and optical schematic.

Synchronized Capture and Alignment

A custom Python-based automation package manages the display and camera triggers. Precise delays (200ms between cameras and 500ms between image changes) are implemented to ensure sensor stability and eliminate motion blur.

Because physical alignment is never perfect, the team used the Kornia library to computationally align the reconstructed lensless images to the ground truth. This process learns a homography mapping that ensures pixel-to-pixel correspondence. For real-time calibration and data validation, the Fast Iterative Shrinkage-Thresholding Algorithm (FISTA) was used for reconstruction.

The MIRFLICKR-25000 Lensless Dataset

Using the MIRFLICKR-25000 dataset as source material, the authors produced an open-access dataset containing 25,000 images captured across two distinct lensless systems:

  1. DiffuserCam: A system using a Gaussian diffuser.
  2. RML: A system using a Random Multi-focal Lenslet phase mask.

Each entry includes the lensless measurement paired with an undistorted ground truth image from the lensed camera.

Key Insights and Results

The parallel acquisition approach ensures that performance differences between systems are attributable to the optical masks themselves rather than environmental variables like lighting or monitor flicker.

Point Spread Function (PSF) and Fidelity

Analysis of the Point Spread Function (PSF)—the optical "fingerprint" of the mask—revealed that the RML mask produces a sharper main lobe and lower sidelobes compared to standard diffusers. This results in a higher Signal-to-Noise Ratio (SNR) and improved resolution in the final reconstructions.

Visual comparison of reconstructed images from DiffuserCam and RML systems against the ground truth.

Bridging the Sim-to-Real Gap

A major challenge in lensless imaging is the "sim-to-real gap," where models trained on simulated data fail on real hardware due to unmodeled non-idealities like stray light or sensor noise. By providing 25,000 real experimental images, this pipeline enables the training of high-capacity models, such as transformers, which were previously limited by smaller datasets.

Conclusion and Future Directions

The authors have open-sourced their hardware designs and synchronization code to foster community adoption. While the current setup uses standard LCD monitors, future iterations could benefit from OLED displays to improve contrast and black levels. Additionally, the researchers plan to extend the pipeline to capture 3D datasets, microscopy-scale data, and video, which would facilitate the development of space-time reconstruction algorithms.

Summary updated

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.
BibTeX
@misc{hung2025scalable,
  author = {Clara S. Hung and Leyla Kabuli and V. A. Ponomarenko and Laura Waller},
  title = {{Scalable dataset acquisition for data-driven lensless imaging}},
  year = {2025},
  eprint = {2501.13334},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.13334},
}
Figure