Research map / Image generation / Autoregressive and latent generation
OpticalNet: An Optical Imaging Dataset and Benchmark Beyond the Diffraction Limit
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces OpticalNet, a building-block optical imaging dataset of FIB-fabricated subwavelength square units (simulated and experimental) and benchmarks CNN and transformer models on translating diffraction images to object images beyond the diffraction limit.
Problem
Resolving subwavelength objects with conventional visible-light microscopy is limited by diffraction, and there is no general open dataset for learning diffraction-image-to-object-image mappings beyond the diffraction limit.
Approach
Decompose arbitrary-shaped objects into small subwavelength square units (building blocks), train a pixel-level image-to-image model on the blocks, and generalize to complex unseen shapes (Light calligraphy, Siemens star) by stitching overlapping block predictions.
Key results
- On simulated Block data, the transformer reached 84.77% accuracy and 79.51 F1 on Block test, and 69.43% accuracy on the Siemens Star test set.
- On experimental data (3x3 GT), the transformer achieved 80.31% accuracy on Block and 55.81% on Siemens Star, outperforming ResUNet-a, AttU-Net, ResNet-18 and ResNet-34.
- Authors report modest Siemens Star performance (about 50% accuracy) on experimental data, indicating room for improvement.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use transformer instead of CNN-based methods | OpticalNet models trained on simulation Block data · Block, Light, and SS | overall performance | Helped |
| Use ResNet-18 instead of transformer | OpticalNet models trained on simulation Block data · Light | F1: 76.47 vs 77.62 | Helped |
| Train on realistic experimental data rather than simulation data | OpticalNet vision models · Block, Light, and SS | performance | Mixed |
| Vary GT block dimension from 3×3 to 5×5 or 7×7 (ablation) | transformer trained on experimental Block data · Block, Light, and SS | accuracy: 80.31, 80.17, 79.95; 74.71, 77.98, 78.11; 55.81, 53.50, 52.70 | No clear effect |
| Increase GT block dimension (ablation) | transformer trained on experimental Block data · SS stitched predictions | visual quality | Mixed |
| Use transformer rather than ResNet-34 | models trained on experimental OpticalNet data · Block, Light, and SS | prediction quality | Helped |
Surprising findings
- Transformers outperform CNN-based models mainly on noisy experimental data, while ResNet variants match them in simulation.
- Increasing ground-truth block size improves visual quality of stitched predictions but adds noise, especially far from the Siemens Star center.
Limitations
- Modest performance on the Siemens Star test set (about 50% accuracy) with blurry or noisy predictions.
- Pixelated (discrete) imaging only; continuous imaging not yet achieved.
- Sim-to-real transfer not yet explored; experiments limited to a single 633 nm setup.
Experimental setup
Datasets: OpticalNet Block dataset (simulated and experimental) OpticalNet Light test set OpticalNet Siemens Star test set
Benchmarks: OpticalNet Block OpticalNet Light OpticalNet Siemens Star
Metrics: accuracy F1-score Jaccard index
Baselines: ResUNet-a AttU-Net ResNet-18 ResNet-34
Builds on: U-Net Attention U-Net ResNet Vision Transformer
Closest papers by meaning
- V2V3D: View-to-View Denoised 3D Reconstruction for Light Field Microscopy (CVPR 2025 · similarity 0.85)
- All-Optical Nonlinear Diffractive Deep Network for Ultrafast Image Denoising (CVPR 2025 · similarity 0.84)
- Single View Refractive Index Tomography with Neural Fields (CVPR 2024 · similarity 0.84)
- Subsurface Scattering for Gaussian Splatting (NeurIPS 2024 · similarity 0.84)
- Continuous Optical Zooming: A Benchmark for Arbitrary-Scale Image Super-Resolution in Real World (CVPR 2024 · similarity 0.83)
- OMG: Opacity Matters in Material Modeling with Gaussian Splatting (ICLR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.