🌧️ ClearView: Neural Image Deraining

Fast, deployement-ready static image deraining model for autonomous driving, surveillance, and photo restoration.
30.9 PSNR / 0.914 SSIM on Rain1400 • ~15ms inference (RTX 4070) • L1 loss + vanilla UNet = best results

---

🚀 Quick Start

Try Online

👉 Live Demo on HuggingFace

Install & Run Locally

git clone https://github.com/dronefreak/clearview.git
cd clearview
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt && pip install -e .

Inference

clearview-infer --image rainy.jpg --weights clearview-unet.pth --output clean.jpg

Train

clearview-train \
  --data-dir /path/to/Rain1400 \
  --model unet --loss l1 --optimizer adamw --lr 1e-4 \
  --batch-size 24 --epochs 200 --dataset-type rain1400

---

📊 Performance

Model	PSNR	SSIM	Params	Speed
UNet (L1)	30.91	0.914	7.8M	~15ms
Attention UNet	30.04	0.910	8.9M	~20ms

✅ Key insight: L1 loss alone outperforms complex multi-loss combos.
⚠️ Limitations: Trained on synthetic rain; slight texture smoothing.

→

---

🏗️ Architecture Highlights

• Backbone: Vanilla UNet (4 encoder/decoder blocks + skip connections)
• Output: Sigmoid-bounded to [0,1]
• Loss: Pixel-wise L1 (loss = L1(pred, target))
• Why not attention? No measurable gain—adds latency and params.

---

📦 Pretrained Model

Download from Hugging Face:

from huggingface_hub import hf_hub_download
weights = hf_hub_download("dronefreak/clearview-unet", "clearview-unet.pth")

🔗 Model Card

---

📚 Supported Datasets

• Rain1400 (recommended): 12.6K train / 1.4K test
• Rain100H/L: Heavy/light rain variants
• Custom: Organize as train/{rainy_image,ground_truth}

---

🛠️ Advanced Usage

• Video: Frame-by-frame processing via scripts/video_demo.py (no temporal consistency yet)
• Metrics: clearview-eval reports PSNR, SSIM, MAE, MSE
• Training Tips:
  • Use mixed precision + gradient clipping
  • Early stopping (patience=50)
  • Avoid multi-component losses

---

🔮 Roadmap

• Temporal consistency for video
• Real-world rain dataset
• Mobile deployment (ONNX/TensorRT)
• Snow/fog/haze removal

---

🤝 Contribute

PRs welcome! See CONTRIBUTING.md.
Need help? Open an Issue.

---

📖 Citation

@software{saksena2025clearview,
  author = {Saksena, Saumya Kumaar},
  title = {ClearView: Practical Image Deraining with PyTorch},
  year = {2025},
  url = {https://github.com/dronefreak/clearview}
}

License: Apache 2.0
Author: Saumya Kumaar Saksena (@dronefreak)

---