| # Learnable Telegraph Diffusion for Image Denoising |
|
|
| This repository contains stage-wise trained denoising models for additive Gaussian noise removal, together with a TNRD-style baseline and a classical PDE baseline. |
|
|
| The results below summarize the latest complete overnight sweep: |
|
|
| - log directory: `logs/overnight_20260408_013608` |
| - evaluated datasets: `Set12`, `BSD68` |
| - noise levels: `sigma = 15, 25, 50, 75` |
| - stage count: `5` |
|
|
| Note: the earlier run in `logs/overnight_20260408_013234` was a scheduler-bug run and is not used for the summary below. |
|
|
| ## Main Takeaways |
|
|
| - The strongest model in this sweep is the `Finetuned TNRD baseline` on both `Set12` and `BSD68` at every tested noise level. |
| - For both MLP and RBF parameterizations, the `No-wave` variant outperformed the `Telegraph` variant throughout this sweep. |
| - The `RBF` parameterization consistently outperformed the corresponding `MLP` parameterization. |
| - End-to-end fine-tuning improved every model family over its stage-wise checkpoint. |
| - All learned models clearly outperformed the classical PDE baseline at the tested noise levels. |
|
|
| ## Best Results |
|
|
| | Dataset | Sigma | Best model | PSNR (dB) | |
| |---|---:|---|---:| |
| | BSD68 | 15 | Finetuned TNRD baseline | 30.90 | |
| | BSD68 | 25 | Finetuned TNRD baseline | 28.36 | |
| | BSD68 | 50 | Finetuned TNRD baseline | 25.43 | |
| | BSD68 | 75 | Finetuned TNRD baseline | 23.91 | |
| | Set12 | 15 | Finetuned TNRD baseline | 31.85 | |
| | Set12 | 25 | Finetuned TNRD baseline | 29.33 | |
| | Set12 | 50 | Finetuned TNRD baseline | 26.05 | |
| | Set12 | 75 | Finetuned TNRD baseline | 24.18 | |
|
|
| ## Plots |
|
|
| ### Base Models |
|
|
|  |
|
|
| ### Finetuned Models |
|
|
|  |
|
|
| ## BSD68 Results |
|
|
| ### Base Models |
|
|
| | Method | 15 | 25 | 50 | 75 | |
| |---|---:|---:|---:|---:| |
| | MLP Telegraph | 28.31 | 25.06 | 22.86 | 20.88 | |
| | MLP No-wave | 29.08 | 26.87 | 23.87 | 21.71 | |
| | RBF Telegraph | 27.97 | 25.52 | 22.70 | 20.89 | |
| | RBF No-wave | 30.46 | 27.86 | 24.47 | 22.30 | |
| | TNRD baseline | 30.41 | 27.85 | 24.58 | 22.42 | |
|
|
| ### Finetuned Models |
|
|
| | Method | 15 | 25 | 50 | 75 | |
| |---|---:|---:|---:|---:| |
| | Finetuned MLP Telegraph | 29.90 | 27.30 | 24.42 | 22.40 | |
| | Finetuned MLP No-wave | 29.88 | 27.61 | 24.60 | 22.93 | |
| | Finetuned RBF Telegraph | 30.56 | 27.70 | 24.65 | 23.26 | |
| | Finetuned RBF No-wave | 30.79 | 28.30 | 25.23 | 23.74 | |
| | Finetuned TNRD baseline | 30.90 | 28.36 | 25.43 | 23.91 | |
|
|
| ## Set12 Results |
|
|
| ### Base Models |
|
|
| | Method | 15 | 25 | 50 | 75 | |
| |---|---:|---:|---:|---:| |
| | MLP Telegraph | 29.19 | 25.92 | 23.32 | 20.99 | |
| | MLP No-wave | 29.59 | 27.47 | 24.54 | 22.28 | |
| | RBF Telegraph | 29.19 | 26.50 | 23.32 | 21.36 | |
| | RBF No-wave | 31.45 | 28.97 | 25.43 | 23.12 | |
| | TNRD baseline | 31.43 | 28.94 | 25.54 | 23.20 | |
|
|
| ### Finetuned Models |
|
|
| | Method | 15 | 25 | 50 | 75 | |
| |---|---:|---:|---:|---:| |
| | Finetuned MLP Telegraph | 30.63 | 28.03 | 24.76 | 22.38 | |
| | Finetuned MLP No-wave | 30.64 | 28.40 | 25.10 | 23.12 | |
| | Finetuned RBF Telegraph | 31.51 | 28.66 | 25.23 | 23.51 | |
| | Finetuned RBF No-wave | 31.74 | 29.23 | 25.92 | 24.05 | |
| | Finetuned TNRD baseline | 31.85 | 29.33 | 26.05 | 24.18 | |
|
|
| ## Classical PDE Baseline |
|
|
| The latest overnight run evaluated the classical PDE baseline at `sigma = 15, 50, 75`. |
|
|
| | Dataset | 15 | 25 | 50 | 75 | |
| |---|---:|---:|---:|---:| |
| | BSD68 | 26.00 | - | 19.56 | 15.08 | |
| | Set12 | 26.73 | - | 19.55 | 14.98 | |
|
|
| ## Notes |
|
|
| - The learned models were trained stage-wise first, then optionally fine-tuned end-to-end on the same noise level. |
| - Fine-tuned checkpoints and base checkpoints were both evaluated using the same sigma-specific setup. |
| - `evaluate_checkpoints.py` produced the main result table used here. |
| - `plot_experiment_results.py` generated the plots in `plots/`. |
|
|
