Implementation and evaluation of classical and deep learning-based image denoising methods. The project compares Non-Local Means (NLM), BM3D, and U-Net under different noise models (Gaussian, Salt & Pepper), using PSNR as the primary performance metric.
This project presents a comparative experimental study between classical statistical denoising algorithms and deep learning-based architectures for image restoration.
Three approaches were implemented and evaluated:
- Non-Local Means (NLM)
- BM3D
- U-Net Denoising Autoencoder
The evaluation was conducted on benchmark datasets (CBSD68 and Kodak24) under Gaussian and Salt & Pepper noise models. Performance was measured using Peak Signal-to-Noise Ratio (PSNR).
The experimental findings demonstrate that performance superiority is strongly dependent on noise distribution assumptions and model inductive bias.
- Training Dataset: CBSD68
- Evaluation Dataset: Kodak24 (24 high-quality color images)
- Resolution: 768×512 / 512×768
PSNR = 10 × log10(MAX² / MSE)
Where MAX = 255 for 8-bit images.
Higher PSNR values indicate better reconstruction quality.
Interpretation:
- PSNR > 30 dB → Very high quality
- 20–30 dB → Acceptable quality
- < 20 dB → Low quality
- Gaussian Noise: μ = 0, varying σ
- Salt & Pepper Noise: 5–15% corruption
- U-Net training:
- Optimizer: Adam
- Learning rate: 0.001
- Epochs: 50–100
- Loss: MSE
- Batch size: 16
- Noise injection (Gaussian / Salt & Pepper)
- Denoising via:
- NLM
- BM3D
- U-Net
- PSNR computation
- Aggregated benchmarking on Kodak24
- Comparative analysis
The pipeline enables controlled experimentation, modular implementation, and reproducible benchmarking across noise configurations.
| Method | PSNR (dB) | Improvement |
|---|---|---|
| Noisy Image | 20.22 | - |
| NLM | 27.62 | +7.40 dB |
| BM3D | 29.94 | +9.72 dB |
| U-Net | 26.70 | +6.48 dB |
Observation
- BM3D achieved the highest PSNR (29.94 dB)
- Classical collaborative filtering remains superior for Gaussian noise
- U-Net underperformed compared to BM3D in this configuration
| Method | PSNR (dB) | Improvement |
|---|---|---|
| Noisy Image | 18.25 | - |
| NLM | 18.30 | +0.05 dB |
| BM3D | 23.65 | +5.40 dB |
| U-Net | 23.83 | +5.58 dB |
Observation
- U-Net achieved the best performance (23.83 dB)
- Deep learning methods proved more robust to impulse noise
- NLM failed to handle Salt & Pepper noise effectively
BM3D preserves structural details and edges more effectively under Gaussian noise, while U-Net provides competitive smoothing but slightly lower PSNR.
U-Net demonstrates superior robustness to impulse noise, removing sparse corruption while maintaining structural consistency.
The bar plot summarizes the quantitative benchmarking across noise types.
| Method | Gaussian Rank | Salt & Pepper Rank |
|---|---|---|
| BM3D | 🥇 1st | 2nd |
| U-Net | 3rd | 🥇 1st |
| NLM | 2nd | 3rd |
- BM3D outperforms deep learning models under Gaussian noise (+3.24 dB over U-Net).
- U-Net generalizes better across different noise distributions.
- NLM performs adequately only under Gaussian noise.
- Model superiority strongly depends on noise distribution assumptions.
- Classical methods exploit strong inductive priors, while neural networks provide flexibility.
This project was designed and evaluated as an end-to-end denoising pipeline, focusing on:
- Reproducible experimentation
- Modular architecture
- Quantitative benchmarking
- Hyperparameter tuning
- Performance-driven comparison
The objective was structured evaluation under controlled experimental conditions rather than isolated model testing.
- Classical methods require prior noise estimation.
- U-Net provides faster inference after training.
- GPU acceleration significantly reduces training time.
- The architecture can be extended to real-time or video denoising applications.
- Neural models can be exported (e.g., ONNX) for optimized deployment.
- Limited training data for U-Net
- Evaluation based solely on PSNR (no SSIM or perceptual metrics)
- Fixed noise levels
- No real-world noisy image evaluation
- Add SSIM and perceptual metrics
- Deploy via REST API (FastAPI)
- Convert trained model to ONNX
- Implement blind/mixed-noise denoising
- Extend to video denoising
image-denoising-benchmark/
│
├── notebooks/
│ └── DENOISING_PROJECT.ipynb
│
├── report/
│ └── Image_Denoising_Report.pdf
│
├── slides/
│ └── Presentation.pptx
│
├── results/
│ ├── kodak1_gaussian_comparison.png
│ ├── kodak1_saltpepper_comparison.png
│ └── psnr_plot.png
│
├── requirements.txt
└── README.md
git clone https://github.com/KonstantinosBountou/image-denoising-benchmark.git
cd image-denoising-benchmark
pip install -r requirements.txtjupyter notebook notebooks/DENOISING_PROJECT.ipynbMIT License