Research map / Label-efficient and robust learning / Knowledge distillation
Unsupervised Deep Learning for Phase Retrieval via Teacher-Student Distillation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use teacher-student distillation instead of testing with the teacher model alone (ablation) | Unsupervised PR model · Uniform CDPs, J=1, γ=9, prDeep12 | PSNR (dB): 35.21 -> 35.94 | Helped |
| Use the noise-resistant self-expressive loss LT rather than the measurement-fitting loss (ablation) | Teacher-student PR model · Uniform CDPs, J=1, γ=9, prDeep12 | PSNR (dB): 35.22 -> 35.94 | Helped |
| Add noise injection when preparing student training data (ablation) | Teacher-student PR model · Uniform CDPs, J=1, γ=9, prDeep12 | PSNR (dB): 35.62 -> 35.94 | Helped |
| Reuse intermediate teacher estimates for student training (ablation) | Teacher-student PR model · Uniform CDPs, J=1, γ=9, prDeep12 | PSNR (dB): 35.76 -> 35.94 | Helped |
| Add image data augmentation to student training data preparation (ablation) | Teacher-student PR model · Uniform CDPs, J=4, γ=9, prDeep12 | PSNR (dB): 40.81 -> 41.09 | Helped |
| Add condition-aware blocks (ablation) | Unfolding PR network · Uniform CDPs, J=4, γ=9, prDeep12 | PSNR (dB): 40.66 -> 41.09 | Helped |
| Use ConvLSTM cells for cross-stage feature delivery instead of convolutional layers (ablation) | Unfolding PR network · Uniform CDPs, J=4, γ=9, prDeep12 | PSNR (dB): 40.55 -> 41.09 | Helped |
| Train without paired ground-truth data, using the proposed unsupervised scheme rather than supervised MSE (ablation) | Teacher-student PR model · Uniform CDPs, J=4, γ=9, prDeep12 | PSNR (dB): 41.02 -> 41.09 | No clear effect |
| Use the proposed unsupervised method rather than DPUNet on compressive uniform CDPs | PR models · Compressive uniform CDPs, ρ=0.50, prDeep12 | PSNR (dB): 33.18 -> 33.19; 28.63 -> 28.86 | Mixed |
| Use the proposed unsupervised method rather than the strongest listed baseline | PR models · Holographic and ptychographic patterns | PSNR (dB): 33.13 -> 33.81; 29.58 -> 30.28; 25.12 -> 25.68; 23.30 -> 24.08 | Helped |
| Apply the method with mismatched rather than matched noise statistics | Unsupervised PR model · CDPs, J=4, prDeep12 | PSNR (dB): 28.61 -> 28.12 | Hurt |
Closest papers by meaning
- Unsupervised Deep Unrolling Networks for Phase Unwrapping (CVPR 2024 · similarity 0.88)
- Untrained Neural Nets for Snapshot Compressive Imaging: Theory and Algorithms (NeurIPS 2024 · similarity 0.86)
- Using Powerful Prior Knowledge of Diffusion Model in Deep Unfolding Networks for Image Compressive Sensing (CVPR 2025 · similarity 0.84)
- A Fast and Provable Algorithm for Sparse Phase Retrieval (ICLR 2024 · similarity 0.84)
- A Unified Framework for Forward and Inverse Problems in Subsurface Imaging using Latent Space Translations (ICLR 2025 · similarity 0.84)
- UGoDIT: Unsupervised Group Deep Image Prior Via Transferable Weights (NeurIPS 2025 · similarity 0.84)
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.