ConvRML: High-Quality Lensless Imaging with Random Multi-Focal Lenslets

Technical Summary of ConvRML: High-Quality Lensless Imaging with Random Multi-Focal Lenslets

Problem Statement
Mask-based lensless imaging offers a pathway to compact, low-cost, and compressive cameras by replacing traditional lenses with thin phase masks. However, these systems historically suffer from poor reconstruction quality due to ill-posed inverse problems, model mismatches, and the limitations of compressive encoding. A primary bottleneck is the trade-off in optical encoder design: high-multiplexing encoders (like diffusers) provide redundancy necessary for compressed sensing but reduce measurement contrast, particularly in dense natural scenes. Conversely, low-multiplexing designs often lack the precision required for high-quality reconstruction. Furthermore, algorithmic development has largely proceeded independently of optical design, with deep learning architectures often evaluated on datasets generated by high-multiplexing encoders, leaving a gap in understanding how to best pair modern algorithms with improved, lower-multiplexing optical measurements.

Methodology
The authors present ConvRML, a system that integrates advances in both optical hardware and computational reconstruction:

1. Optical Encoder (RML): The system utilizes a Random Multi-Focal Lenslet (RML) phase mask. Unlike traditional diffusers or regular microlens arrays, the RML consists of overlapping lenslets with varying focal lengths placed at pseudo-random positions. This design minimizes correlations between shifted copies of the Point Spread Function (PSF), reduces multiplexing to preserve measurement contrast, and maintains a wide field of view (FOV). The mask is fabricated using high-resolution two-photon lithography, ensuring precision and reproducibility. The multi-focal nature of the lenslets provides an extended depth of focus, reducing sensitivity to axial misalignment.
2. Reconstruction Architecture: The authors implement a ConvNeXt-based deep learning architecture. This convolutional neural network is chosen for its ability to achieve large effective receptive fields through cascaded layers, enabling it to handle the long-range dependencies inherent in multiplexed measurements without the computational overhead or specific architectural constraints of attention-based models (e.g., Transformers).
3. Standardized Evaluation Framework: To isolate the effects of optical encoding from algorithmic performance, the authors constructed a parallel imaging setup. This setup simultaneously captures scenes using three systems under identical conditions: the RML lensless imager, a standard diffuser lensless imager, and a conventional lens system (serving as ground truth).
4. Dataset Contribution: The authors generated and open-sourced the Parallel Lensless Dataset (PLD), containing 100,000 measurements for each system (RML, diffuser, and lens). This dataset is significantly larger than previous open-source benchmarks (e.g., DiffuserCam's 25,000 or PhlatCam's 10,000) and supports controlled comparisons across different optical encoders.

Key Results

• Optical Performance: Quantitative analysis using Modulation Transfer Function (MTF) and Mutual Information (MI) demonstrates that the RML outperforms the diffuser. The RML preserves higher spatial frequencies and maintains significantly more information in the presence of read noise and limited sensor bit depth (e.g., 4-bit quantization). While the diffuser signal often collapses into a single quantization bin under noise, the RML retains distinguishable features.
• Reconstruction Performance:
  • Optical Impact: When paired with the same ConvNeXt architecture, the RML system achieves a 2.99 dB improvement in PSNR and a 0.116 increase in SSIM over the diffuser system, confirming that improved optical measurements directly translate to higher-quality reconstructions.
  • Architectural Impact: On the high-multiplexing DiffuserCam benchmark, ConvNeXt achieved 25.37 dB PSNR, outperforming state-of-the-art attention-based architectures (Lensless Imaging Transformer, Swin Transformer, Vision Transformer) and the U-Net baseline. Notably, ConvNeXt achieved this with moderate model capacity (36M parameters), whereas larger attention-based models (e.g., Swin Transformer with 59.4M parameters) performed worse.
  • Data Efficiency: Training ConvNeXt on the RML dataset showed performance gains up to 50,000 measurements, after which returns diminished. This suggests the architecture is well-matched to this dataset size, whereas larger attention-based models did not necessarily benefit from increased capacity or data in the same way.
• Generalization: The trained ConvNeXt models successfully generalized to reconstruct physical real-world objects (varying materials like fabric, metal, and plastic) placed in front of the lensless systems, recovering fine textures and sharp edges that the diffuser system failed to resolve.

Significance and Claims
The paper claims that the ConvRML system establishes a new standard for high-quality, compact lensless imaging by demonstrating that improved measurement quality paired with powerful, appropriately matched algorithms yields maximum reconstruction performance.

Key contributions include:

1. Hardware: A precision-manufactured RML phase mask that offers reduced multiplexing and improved contrast compared to diffusers, fabricated with high reproducibility.
2. Software: A ConvNeXt reconstruction architecture that outperforms attention-based models, suggesting that carefully designed convolutional networks are sufficient for handling long-range dependencies in lensless imaging without the complexity of Transformers.
3. Resources: The release of the Parallel Lensless Dataset (PLD), the largest open-source experimental lensless imaging dataset to date, which facilitates standardized evaluation and future development of reconstruction architectures.

The authors conclude that their work provides the necessary resources and evidence to support the continued development of high-quality, compressive lensless imagers, moving beyond the limitations of previous high-multiplexing, low-precision systems.