🐕 DetectoDog: Dog Breed Classification

DetectoDog is a deep learning project that classifies dog breeds from images using transfer learning. The primary goal is to power a mobile application capable of identifying dog breeds from user-submitted photos.

📱 Project Goal

The final deliverable for this project is a mobile application that:

• Accepts an image of a dog
• Runs an optimised on-device model
• Displays the predicted breed with confidence score
• Optionally provides breed-related information

The training in this notebook prepares the model for deployment on resource-constrained environments (e.g. phones or tablets), using lightweight architectures and efficient inference strategies.

---

🧠 Model Training (Notebook Overview)

This notebook covers the training process for the DetectoDog classifier:

• ✅ Dataset download and preparation (Stanford Dogs Dataset)
• ✅ Data exploration and augmentation
• ✅ Transfer learning with ResNet50, MobileNetV2, and EfficientNet
• ✅ Accuracy evaluation and performance visualisation
• ✅ Exporting the final trained model

---

🔍 Dataset Summary

• Dataset: Stanford Dogs Dataset
• Classes: 120 breeds
• Images: 20,580 total
• Challenges: Varying image quality, backgrounds, and angles

---

📦 Dependencies

import torch
import torchvision
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
from PIL import Image
from sklearn.metrics import classification_report

Also uses: tqdm, yaml, tarfile, os, scipy.io, etc.

---

🛠 Notebook Features

• Visualisation of class distributions
• Model training and validation curves
• Confusion matrix and metrics
• Sample inferences on unseen images
• Saving the best-performing model for export

---

📲 Mobile Deployment (Next Phase)

The next step will involve:

• Converting the model to TorchScript or ONNX
• Reducing size via pruning or quantisation
• Integrating the model into a mobile app via React Native, Flutter, or Android Studio (Java/Kotlin)
• Performing inference using a mobile-optimised runtime like PyTorch Mobile or TensorFlow Lite

---

🔮 Future Work

• Ensemble modelling to improve accuracy
• More advanced augmentations
• Real-time inference benchmarks
• Breed information popups in app

---

📁 Files

• detectodog_final_structured.ipynb: Full training notebook
• expiriments/models/model.pt: Exported model (after training)
• README.md: Project summary and roadmap