Computational Imaging Project — Motion Deblur & Denoising

Overview

This project addresses an inverse problem in computational imaging: reconstructing high-quality images from degraded observations affected by motion blur and noise.

We compare classical, deep learning, and generative approaches under the same experimental conditions, with the goal of analyzing their strengths and limitations.

---

Task Definition

Given a clean image ( x ), we generate a degraded observation ( y ):

[ y = Kx + n ]

where:

• ( K ): motion blur operator
• ( n ): additive Gaussian noise

Degradation parameters:

• Kernel size: 9
• Motion angle: (chosen experimentally)
• Noise levels: 0.005, 0.01, 0.05, 0.1

All methods are evaluated on the same degraded inputs.

---

Dataset

Dataset used: FFHQ 256x256

🔗 https://huggingface.co/datasets/bitmind/ffhq-256

Preprocessing:

• Resize to 256×256
• Normalize pixel values to [0,1]
• Train / Validation / Test split
• Synthetic degradation generation

All preprocessing choices are documented and reproducible.

---

Methods

We implement and compare four methodological families:

1. Variational Method

• Total p-Variation (TpV)
• ( p \in (0.1, 0.5) )
• Optimization-based reconstruction

---

2. End-to-End Deep Learning

• Architecture: (e.g. UNet / NAFNet / ViT)
• Supervised training on degraded → clean pairs
• Loss functions: L1 / L2 / perceptual (optional)

---

3. Generative Method

• Diffusion Posterior Sampling (DPS)
• Adapted for inverse problems
• Combines diffusion priors with data consistency

---

4. Hybrid Method

• Plug-and-Play (PnP) with HQS
• Iterative reconstruction:
  • Data fidelity step
  • Denoising step using pretrained model

---

Evaluation

We compare all methods using:

Quantitative metrics:

• PSNR
• SSIM

Qualitative analysis:

• Visual comparison of reconstructed images
• Artifact inspection
• Robustness to noise levels

---

Results

• Comparative plots across all methods
• Performance vs noise level
• Trade-offs:
  • accuracy
  • speed
  • stability

---

Project Structure

├── data/ # dataset and preprocessing
├── degradation/ # blur + noise simulation
├── variational/ # TpV implementation
├── models/ # deep learning architectures
├── diffusion/ # DPS implementation
├── pnp/ # Plug-and-Play HQS
├── evaluation/ # PSNR, SSIM, plots
├── notebooks/ # experiments and visualization
├── utils/ # helper functions
└── main.py

🚀 How to Run

1. Clone repository

git clone https://github.com/your-username/your-repo-name.git
cd your-repo-name

2. Install dependencies

pip install -r requirements.txt

3. Download dataset

Follow instructions from: https://huggingface.co/datasets/bitmind/ffhq-256

4. Run experiments

python main.py --method unet python main.py --method tpv python main.py --method dps python main.py --method pnp

Parameter Selection

Parameters are chosen heuristically, including:

Regularization weight (TpV) Network architecture and training setup Number of diffusion steps (DPS) HQS iteration parameters

A detailed discussion is provided in the report and presentation.