Dog Breed Recognition

This project is a web application that classifies dog breeds from images uploaded by users. The application features a React frontend, a Flask backend, and uses a deep learning model to predict the dog breed. It is containerized using Docker and managed with Docker Compose for easy deployment.

Features

• Image Upload: Allows users to upload an image of a dog for classification.
• Breed Prediction: Uses a deep learning model to predict the breed of the dog in the uploaded image.
• Responsive UI: The frontend is built with React, providing an interactive and user-friendly experience.
• Containerization: The entire application is containerized using Docker and orchestrated with Docker Compose, simplifying deployment and scaling.

Technologies Used

• Frontend: React
• Backend: Flask
• Machine Learning: TensorFlow, Keras (ConvNeXt architecture)
• Containerization: Docker, Docker Compose
• Model: Pre-trained ConvNeXt with custom fine-tuning for dog breed classification

Prerequisites

• Docker: Make sure Docker and Docker Compose are installed on your machine. You can download Docker from here.

Getting Started

1. Clone the repository:

   git clone https://github.com/Mateorovere/Dog_breed_classification.git
   cd Dog_breed_classification

2. Build and run the containers using Docker Compose:

   docker-compose up --build

3. Open your browser and navigate to:

   http://localhost:3000
   

Usage

1. Access the application via your browser.
2. Upload an image of a dog.
3. The application will return the predicted dog breed based on the image.

Model Training

The model was trained using the Stanford Dogs Dataset, employing a ConvNeXt-based architecture with transfer learning from ImageNet weights. The training process used data augmentation and early stopping to achieve optimal performance.

To retrain the model, you can modify model_training.py and execute it locally with your own dataset or fine-tune it as needed.

Model Details:

• Architecture: ConvNeXtSmall, fine-tuned with a custom dense layer and dropout for regularization.
• Optimizer: AdamW with a learning rate of 1e-5.
• Loss Function: Categorical Crossentropy.
• Metrics: Accuracy.
• Callbacks: Early stopping, ReduceLROnPlateau, ModelCheckpoint.

The final model is saved as best_model.keras, and class names are saved in class_names.txt.

Acknowledgements

• Stanford Dogs Dataset: Kaggle Link.