Research map / Image generation / Image restoration
All-Optical Nonlinear Diffractive Deep Network for Ultrafast Image Denoising
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.
Proposes N3DNet, a 5-layer all-optical nonlinear diffractive network for image denoising trained with a regularization-assisted DQN, fabricated by 3D printing, and introduces the MIDD mode-image denoising dataset.
Problem
Electronic-chip image denoisers incur latency and high compute cost, while prior all-optical diffractive (D2NN) denoisers lack nonlinearity and noise filtering, so they denoise poorly.
Approach
Adds an image encoding and bandpass pre-denoising module in front of a D2NN, inserts a phase exponential linear (PEL) nonlinear activation after each diffractive layer, and optimizes the phase masks with a DQN variant that penalizes Q-value differences between consecutive steps (RA-DQN).
Key results
- On simulated benchmarks N3DNet beats the best baselines by an average of 0.43 dB PSNR, 0.032 SSIM and 0.066 LPIPS.
- In real experiments on MIDD, N3DNet improves over the best optical baselines by an average 9.15 dB PSNR, 0.248 SSIM and 0.202 LPIPS.
- Computation takes 4.4 µs per 1000×1000 image, about three orders of magnitude faster than electronic methods (16.7 ms), reported as ~3,800x speedup.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use N3DNet instead of the strongest listed baseline | Image denoising · MIDD (σ=20) | PSNR: 29.10 -> 30.15 | Helped |
| Use N3DNet instead of the strongest listed baseline | Image denoising · MIDD (σ=40) | PSNR: 27.25 -> 28.12 | Helped |
| Use N3DNet instead of the strongest listed baseline | Image denoising · CSet9 (σ=25) | PSNR: 31.71 -> 31.68 | Hurt |
| Use N3DNet instead of the strongest listed baseline | Image denoising · CSet9 (σ=50) | PSNR: 29.29 -> 30.01 | Helped |
| Use N3DNet instead of the strongest listed baseline | Image denoising · CBSD68 (σ=25) | PSNR: 30.73 -> 30.70 | Hurt |
| Use N3DNet instead of the strongest listed baseline | Image denoising · CBSD68 (σ=50) | PSNR: 27.68 -> 28.12 | Helped |
| Use N3DNet instead of the strongest listed baseline | Image denoising · Set12 (σ=25) | PSNR: 30.01 -> 30.05 | Helped |
| Use N3DNet instead of the strongest listed baseline | Image denoising · BSD68 (σ=50) | PSNR: 26.83 -> 27.35 | Helped |
| Replace DQN with RA-DQN optimization (ablation) | N3DNet · MIDD (l=50) | convergence loss: 2.70 -> 1.78 | Helped |
| Add the Charbonnier loss term (ablation) | N3DNet loss function · MIDD (l=12 km) and other datasets | PSNR | Helped |
| Add the Fourier loss term to Charbonnier loss (ablation) | N3DNet loss function · MIDD (l=12 km) and other datasets | PSNR | Helped |
| Add the FSIM loss term to Charbonnier loss (ablation) | N3DNet loss function · MIDD (l=12 km) and other datasets | PSNR | Helped |
Surprising findings
- Physical-system results are slightly inferior to simulation results, attributed to 3D printing, optics and encoding errors.
Limitations
- Physical experiments show a performance gap versus simulation due to fabrication and optical system errors.
- Evaluated mainly on mode images from fiber systems; generality to natural images is limited to simulation.
Experimental setup
Datasets: MIDD CSet9 CBSD68 Set12 BSD68 Nam SIDD
Benchmarks: MIDD CSet9 CBSD68 Set12 BSD68 Nam SIDD
Metrics: PSNR SSIM LPIPS latency energy consumption
Baselines: BM3D DnCNN RIDNet MT MPLC ADI PL
Builds on: D2NN DQN Charbonnier loss FSIM loss
Closest papers by meaning
- Optical Diffusion Models for Image Generation (NeurIPS 2024 · similarity 0.85)
- OpticalNet: An Optical Imaging Dataset and Benchmark Beyond the Diffraction Limit (CVPR 2025 · similarity 0.84)
- Neural Tangent Knowledge Distillation for Optical Convolutional Networks (NeurIPS 2025 · similarity 0.84)
- Real-Time Neural Denoising with Render-Aware Knowledge Distillation (AAAI 2025 · similarity 0.84)
- Dual-Rate Diffusion: Accelerating diffusion models with an interleaved heavy-light network (NeurIPS 2026 · similarity 0.84)
- Contrastive Learning for Low-Light Raw Denoising (Student Abstract) (AAAI 2024 · 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.