Dog Breed Classification – ImageNet

End-to-end pipeline for dog breed classification using deep learning in PyTorch, applied to the ImageNet Dogs dataset. This project was developed as part of the study unit CCE3207 – Introduction to Computer Vision under the supervision of Dr. Christian Galea.

---

Project Overview

This project implements a full training and evaluation pipeline for a deep learning-based dog breed classification model. The pipeline includes:

• Exploratory data analysis
• Custom dataset loading with bounding box cropping
• Transfer learning using pretrained CNNs
• Model training, validation, and evaluation
• Final predictions on unseen data

The project is implemented within a single annotated Jupyter notebook, demonstrating practical computer vision techniques and model interpretability strategies.

---

Problem Statement

The task involves classifying dog breeds from images in the ImageNet Dogs subset. Each image is annotated with a bounding box and a breed label. The challenges addressed include:

• Handling class imbalance across 120+ breeds
• Automating image preprocessing and augmentation
• Efficient training using GPU acceleration
• Managing large-scale image datasets using PyTorch Dataset and DataLoader

---

Folder Structure

.
├── Assignment/
│   ├── assignment.ipynb       # Main Jupyter notebook (contains full pipeline)
│   ├── labels.txt             # Mapping of class indices to dog breed labels
│   └── model/                 # Saved model folders (per experiment)
│       ├── resnet18/
│       ├── efficientnet_b0/
│       └── ...
└── README.md

---

Key Features

• Data Exploration: Insights into class distribution, image sizes, and bounding box statistics.
• Custom Dataset Class: Crops and resizes images based on bounding boxes for accurate feature learning.
• Transfer Learning: Fine-tunes pretrained ResNet and EfficientNet backbones on dog breeds.
• Evaluation Metrics: Accuracy, confusion matrix, and top-k classification score analysis.
• Model Selection: Compares multiple CNN backbones and training hyperparameters.

---

Dependencies

The project uses the following Python libraries:

• PyTorch, torchvision
• OpenCV, Pillow, xml
• matplotlib, numpy, os

All code was executed using GPU acceleration on the Networks lab systems provided by the university.

---

How to Run

1. Clone the repository:

git clone https://github.com/GrahamPellegrini/Dog-Breed-Classification.git

2. Open assignment.ipynb in JupyterLab or VSCode.

3. Ensure the dataset is available under a valid path:

/opt/nfs/shared/images/ImageNetDogs

Alternatively, update the notebook to point to a custom dataset path of your choice.

4. Run all notebook cells sequentially to train and evaluate models.

---

Results Snapshot

The following summarizes the performance of all models tested throughout the project:

Model	Accuracy	Loss	Observations
Simple-CNN	5.63%	3.68	Baseline implementation with minimal layers
Improved-CNN	12.41%	4.29	Added depth, batch norm, and bottlenecking
Deep-CNN	42.98%	2.17	Modular 5-layer CNN, stronger generalisation
VGG16-BN	54.27%	2.98	Pretrained model, signs of overfitting
VGG16-1 (Adam+Scheduler)	70.18%	1.53	Frozen conv layers, slower convergence
VGG16-2 (SGD Optimizer)	79.43%	0.69	Best model: stable training, minimal overfitting

Final deployment candidate: VGG16-2 for its accuracy and stable validation trends.

Training logs and saved weights available under model/ directory.

---

---

Author

Graham Pellegrini
B.Eng. (Hons) Computer Engineering
University of Malta
GitHub: @GrahamPellegrini