Research map / Image generation / Image restoration
WaveMo: Learning Wavefront Modulations to See Through Scattering
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.
Learns acquisition-time wavefront modulations (an implicit-neural-representation of 16 Zernike-based patterns) end-to-end with a feedforward U-Net proxy reconstructor, then reuses them with the unsupervised NeuWS algorithm to image through scattering media.
Problem
Imaging through scattering media loses scene frequencies; existing wavefront modulation schemes for phase diversity imaging use random or heuristic patterns, and optimizing modulations through iterative reconstruction is impractical.
Approach
Jointly optimize an MLP-parameterized set of wavefront modulations and a differentiable feedforward proxy reconstruction network end-to-end, so the proxy's reconstruction error guides modulation design; the learned modulations are then decoupled and applied to other reconstruction algorithms such as NeuWS.
Key results
- In simulation, learned modulations raise proxy network PSNR to 30.39 dB vs 26.44 dB (random) and 25.48 dB (none) on 1000 Places365 test images.
- On physical tissue data, the proxy with learned modulations reaches 19.06 PSNR vs 17.57 (random) and 16.53 (none).
- Applying learned modulations to NeuWS improves static-scene PSNR from 11.26 (random) to 14.61 and dynamic-scene PSNR from 8.90 to 12.89.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace no modulations with learned wavefront modulations | Proxy network · MIT Places 365 test set (1,000 images) | PSNR: 25.476 -> 30.391 | Helped |
| Replace random modulations with learned wavefront modulations | Proxy network · MIT Places 365 test set (1,000 images) | PSNR: 26.439 -> 30.391 | Helped |
| Replace no modulations with learned wavefront modulations | Proxy network · MIT Places 365 test set (1,000 images) | SSIM: 0.7640 -> 0.9082 | Helped |
| Replace random modulations with learned wavefront modulations | Proxy network · MIT Places 365 test set (1,000 images) | SSIM: 0.7980 -> 0.9082 | Helped |
| Replace random modulations with learned wavefront modulations | Proxy network · Tissue | PSNR: 17.57 -> 19.06 | Helped |
| Replace random modulations with learned wavefront modulations | Proxy network · Out-of-distribution scenes | PSNR: 9.90 -> 10.71 | Helped |
| Replace random modulations with learned wavefront modulations | NeuWS · Static scenes | PSNR: 11.26 -> 14.61 | Helped |
| Replace random modulations with learned wavefront modulations | NeuWS · Dynamic scenes | PSNR: 8.90 -> 12.89 | Helped |
| Replace random modulations with learned wavefront modulations | Proxy network · Tissue | SSIM: 0.48 -> 0.58 | Helped |
| Replace random modulations with learned wavefront modulations | Proxy network · Out-of-distribution scenes | SSIM: 0.30 -> 0.32 | Helped |
| Replace random modulations with learned wavefront modulations | NeuWS · Static scenes | SSIM: 0.21 -> 0.38 | Helped |
| Replace random modulations with learned wavefront modulations | NeuWS · Dynamic scenes | SSIM: 0.23 -> 0.33 | Helped |
Surprising findings
- The learned modulations were optimized only with an MSE reconstruction loss, yet they implicitly raise the MTF at high spatial frequencies.
Limitations
- Proxy network generalizes poorly to out-of-distribution targets (PSNR ~10.7 on out-of-distribution scenes)
- Validated only on monochromatic imaging; broadband extension left to future work
- Iterative NeuWS needs multiple modulated measurements and cannot reconstruct from a single measurement
Experimental setup
Datasets: MIT Places365 adipose tissue pathological slide images
Benchmarks: MIT Places365 test set (1000 images) physical scattering experiments on adipose tissue and out-of-distribution targets
Metrics: PSNR SSIM MTF
Baselines: No modulation Random modulation NeuWS
Builds on: NeuWS phase diversity imaging end-to-end learning of optical systems
Closest papers by meaning
- Understanding Multi-layered Transmission Matrices (CVPR 2025 · similarity 0.85)
- Decoupling Scattering: Pseudo-Label Guided NeRF for Scenes with Scattering Media (AAAI 2025 · similarity 0.85)
- Virtual Scanning: Unsupervised Non-line-of-sight Imaging from Irregularly Undersampled Transients (NeurIPS 2024 · similarity 0.85)
- RxGS: Receiver-Generalizable 3D Gaussian Splatting for Radio-Frequency Data Synthesis (NeurIPS 2026 · similarity 0.83)
- A Physics-Informed Blur Learning Framework for Imaging Systems (CVPR 2025 · similarity 0.82)
- Subsurface Scattering for Gaussian Splatting (NeurIPS 2024 · similarity 0.82)
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.