Image Segmentation: Classical vs Deep Learning Approaches

---

This project presents a comparative study between two binary semantic segmentation approaches applied to the Oxford-IIIT Pet Dataset:
(1) a classical approach based on OpenCV's GrabCut algorithm, and
(2) a convolutional encoder-decoder architecture implemented in PyTorch.

---

Requirements

• Python 3.8+
• PyTorch
• OpenCV
• torchvision
• matplotlib
• tqdm

Install the dependencies:

pip install -r requirements.txt

---

Setup

Before running any scripts, set the PYTHONPATH to the project root:

export PYTHONPATH=$(pwd)

---

How to Run

1.Train Neural Network (The dataset will be downloaded automatically)

python methods/deep.py --mode train --plot

This will save :

• the model in models/trained/unet.pth
• the training metrics in outputs/training/

---

2. Run U-Net Prediction

python methods/deep.py --mode predict

This will save the segmentation masks in outputs/unet/

---

5. Run GrabCut

python methods/classical.py --all

This will save the segmentation masks in outputs/classical/

---

6.Evaluate Results

python evaluation/evaluate.py

This will show the metrics for each model and save the comparative histograms in:

• outputs/classical_metrics/
• outputs/unet_metrics/

---

7. Visual Comparison

python scripts/visualize.py

Saves the figure in:
outputs/visualize/visualize.pdf

---

Project Structure

datasets/         # Dataset loader
models/           # encoder -decoder model
methods/          # Training and classical segmentation scripts
utils/            # Metrics
evaluation/       # Evaluation pipeline
scripts/          # Visual comparison

---