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

TL;DR

ConvRML addresses the persistent gap between optical encoding and reconstruction quality in lensless imaging by pairing a precision-fabricated random multi-focal lenslet (RML) phase mask with a ConvNeXt-based reconstruction, supported by a standardized parallel dataset of $100000$ measurements per system. The approach yields substantial gains in reconstruction fidelity—up to $6.68$ dB PSNR over attention-based methods—and demonstrates improved high-frequency information transfer via MTFs and MI analyses. The work provides large, open-source datasets and a rigorous, parallel evaluation framework to enable fair comparisons across optical encoders and reconstruction architectures, and shows real-world object reconstructions that generalize beyond display-based training data. Overall, ConvRML demonstrates that improved optical encoding paired with modern convolutional reconstruction can enable high-quality, compact, and compressive lensless cameras with broad practical potential.

Abstract

Mask-based lensless imagers use simple optics and computational reconstruction to design compact form factor cameras with compressive imaging ability. However, these imagers generally suffer from poor reconstruction quality. Here, we describe several advances in both hardware and software that result in improved lensless imaging quality. First, we use a precision-manufactured random multi-focal lenslet (RML) phase mask to produce improved measurements with reduced multiplexing. Next, we implement a ConvNeXt-based reconstruction architecture, which provides up to 6.68 dB improvement in peak signal-to-noise ratio over state-of-the-art attention-based architectures. Finally, we establish a parallel imaging setup that simultaneously images a scene with RML, diffuser and lens systems, with which we collect datasets with 100,000 measurements for each system, to be used for reconstruction model training and evaluation. Using this standardized system, we quantify the improved measurement quality of the RML compared to a diffuser using the modulation transfer function and mutual information. Our ConvRML system benefits from both the optical and the computational developments presented in this work, and our contributions establish resources to support continued development of high-quality, compact, and compressive lensless imagers.

Figures (11)

• Figure 1: ConvRML system overview. (top) Our high-quality lensless imaging system combines a random multi-focal lenslet (RML) phase mask with a ConvNeXt reconstruction architecture. (middle) To facilitate data-driven training of reconstruction architectures and controlled system comparisons between a lens, a diffuser, and our RML, we collect and open-source datasets with 100,000 measurements for each system. (bottom) ConvNeXt models trained on displayed image datasets for the RML and diffuser lensless imagers produce high-quality reconstructions of real-world objects.
• Figure 2: Random multi-focal lenslet (RML) phase masks result in good measurement quality. a) Illustration and b) experimentally-measured point spread functions (PSFs) for different optical encoders with a fixed numerical aperture $\theta$. A lens focuses light to a single focal point, with the field of view (FOV) decreasing for shorter focal lengths. The RML distributes light across a small number of focal points, preserving a wide FOV with minimal diverging light. The diffuser both focuses and scatters light, resulting in a high-multiplexing caustic pattern. PSFs are contrast stretched for visualization. c) Modulation transfer functions (MTFs) for each design, with the lens system serving as an ideal reference. The RML has improved frequency transfer compared to the diffuser, especially at high spatial frequencies. d) Simulated measurements for each imaging system. The high-multiplexing diffuser PSF generally results in lower contrast measurements, while RML measurements preserve larger local intensity variations. e) Cross-sections of each measurement along with dashed lines showing the loss of dynamic range with quantization. f) Mutual information evaluation of the effects of read noise and limited bit depth. The RML maintains higher information than the diffuser as noise increases (top) and retains nonzero information at low bit depths while the diffuser has approximately zero information (bottom).
• Figure 3: Standardized experimental comparison of RML and diffuser systems. a) Parallel dataset acquisition with RML, diffuser, and lens (ground truth) systems under identical imaging conditions facilitates standardized comparisons. b) The ConvNeXt-based reconstruction architecture creates multi-scale feature maps, which are concatenated and progressively compressed. c) Example reconstructions comparing ground truth, our RML with ConvNeXt reconstruction (ConvRML), and a diffuser with ConvNeXt reconstruction. Using an RML achieves higher reconstruction quality than the diffuser, as quantified by the structural similarity index measure (SSIM).
• Figure 4: Comparison of reconstruction models evaluated on the DiffuserCam Lensless Mirflickr Dataset MonakhovaLearning. Our ConvNeXt model provides reconstructions that most closely match ground truth images for this high-multiplexing diffuser system.
• Figure 5: Comparison of reconstruction models evaluated on our Parallel Lensless Dataset. Our ConvRML system provides reconstructions that are best matched to ground truth images for the low-multiplexing RML dataset. ConvNeXt reconstructions for high-multiplexing diffuser measurements are also provided as a reference. Both the RML and the ConvNeXt reconstruction architecture contribute to image quality improvements over state-of-the-art.