This project explores deep learning models for restoring blurry low-resolution images. It compares a custom DnCNN-based super-resolution model with a transformer-style NAFNet architecture. Both were implemented and evaluated on the GOPRO and RealBlur-R datasets using Colab Pro (A100 GPU).

• DnCNN-SR: Residual CNN + PixelShuffle-based upsampling
• NAFNet: Transformer-inspired architecture (implemented, not used for demo)
• Metrics: PSNR, SSIM, LPIPS
• Losses: MSE, perceptual (VGG), LPIPS
• Training: AMP, early stopping, ReduceLROnPlateau
• Output visualization and metric summary

You can quickly run the pretrained DnCNN model using:

demo.ipynb

No training needed — just load the weights and run on your own images.

DnCNN showed the best perceptual and numerical performance. NAFNet was successfully implemented but not used in the final visualization due to training instability.

Left: LR input (padded, 360x640) | Center: 2x SR output (DnCNN, 720x1280) | Right: HR ground truth
PSNR: 26.80 | SSIM: 0.8020 | LPIPS: 0.2313

We originally tried this cascade:

DnCNN → UNet → EDSR

While promising in theory, this chain:

• Suffered from compounding artifacts
• Was harder to converge
• Did not outperform DnCNN alone in PSNR/SSIM/LPIPS

📌 Conclusion: well-designed single models + quality upsampling outperform deep cascades in image restoration.

image-restoration/
├── old_joint_model_code/   # Original full pipeline code archive
├── results/                # Output samples + originals + result visualizations + metrics_results
├── LICENSE
├── README.md               # You're reading it
├── demo.ipynb              # Run DnCNN on test images (quick start)
├── dncnn_sr.ipynb          # Full DnCNN model training + results
├── nafnet.ipynb            # Full NAFNet implementation + training (optional)
└── requirements.txt

• GOPRO_Large Dataset
• RealBlur-R Dataset

You can download pretrained DnCNN weights here: Google Drive

pip install -r requirements.txt

Then launch demo.ipynb to run DnCNN on your own input images.

This repo includes a pretrained model and demo script. You do not need to train anything to test results.

• torch
• torchvision
• lpips
• tqdm
• matplotlib
• scikit-image
• opencv-python

Maintained by Hank Song For questions, feel free to open an issue or reach out.