🧽 U-Net Denoising: Scientific Image Restoration with Deep Learning

🧠 Project Overview

This project implements a lightweight U-Net model to denoise grayscale scientific images corrupted with synthetic Gaussian noise. It mimics real-world noise introduced by microscopes, satellite sensors, or medical scanners, and applies a deep learning surrogate to restore the original signal.

This work is aligned with surrogate modeling under uncertainty, a core research theme at the GOKUL Lab, IIT Hyderabad.

---

📦 Dataset Usage

We used the BSDS500 grayscale dataset:

🔗 https://www2.eecs.berkeley.edu/Research/Projects/CS/vision/bsds/

Dataset Prep:

• All .jpg images from train/, val/, and test/ were flattened into data/raw/
• Gaussian noise was added using add_noise.py
• Noisy images were saved in data/noisy/

⚠️ Dataset files (data/raw/ and data/noisy/) are not included in the repository to keep it lightweight. Use the noise script to regenerate them.

---

🗂️ Project Structure

Gen AI Projects/
├── data/
│   ├── raw/            # Clean grayscale images (not pushed)
│   ├── noisy/          # Noisy versions of images (auto-generated)
├── images/
│   ├── train/
│   ├── val/
│   └── test/
├── models/
│   └── unet.py         # U-Net architecture
├── outputs/
│   ├── model_last.pth  # Trained model weights (ignored)
│   ├── loss_curve.png  # Training loss plot (ignored)
│   └── sample_epoch3.png  # Denoising output (ignored)
├── U-Net Denoising/
│   ├── add_noise.py    # Adds Gaussian noise to raw images
│   └── evaluate.py     # Computes PSNR and SSIM metrics
├── output.png          # Manual visualization (ignored)
├── train.py            # Main training script
├── requirements.txt    # Dependencies
├── README.md           # This file
└── .gitignore          # Ignore rules

---

🧠 Model Architecture

• Type: U-Net (3-level lightweight variant)
• Input/Output: Grayscale image (1 channel)
• Loss Function: MSE
• Optimizer: Adam (lr = 1e-3)
• Training Resolution: 256×256

---

🏋️ Training Summary

Setting	Value
Epochs	3
Batch Size	8
Final Loss	~0.0011
Device	MacBook Air M3

---

📈 Evaluation Results

Metric	Value
PSNR	29.52 dB
SSIM	0.8407

These results were computed over 500+ denoised samples and indicate strong recovery performance.

---

🖼️ Output Placeholders

Original Image	Noisy Input	Denoised Output

Replace the placeholders with your manually noised and cleaned examples.

---

🚀 Run Instructions

1. Install dependencies:

pip install -r requirements.txt

2. Generate noisy images:

python U-Net\ Denoising/add_noise.py

3. Train the model:

python train.py

4. Evaluate performance:

python U-Net\ Denoising/evaluate.py

---

📌 Credits

Developed by [Your Name] as part of a deep learning portfolio project aligned with research interests at GOKUL Lab, IIT Hyderabad.