🛩️ Aircraft-Classifier

Fine-Grained Visual Classification of Aircraft Using CNNs and PyTorch

This project implements a deep learning solution for fine-grained aircraft classification using Convolutional Neural Networks (CNNs) and PyTorch. The model can accurately classify aircraft images into 10 different aircraft variants using transfer learning with a pre-trained ResNet-18 architecture.

🎯 Project Description

The goal of this project is to use deep learning to classify aircraft photos into fine-grained categories. The FGVC-Aircraft Benchmark dataset, which comprises 102 airplane variants, is used to train a Convolutional Neural Network (CNN). This task offers a realistic and demanding environment for image classification because of the high visual similarity between classes (e.g., different Boeing 737 variants).

Since the dataset contains 10,200 images of aircraft, with 100 images for each of 102 different aircraft model variants, we use a subset of 10 randomly selected classes for this implementation. The objective is to create a deep learning model using PyTorch that can correctly classify aircraft into predetermined categories.

📊 Dataset

FGVC-Aircraft Dataset

Source: University of Oxford Visual Geometry Group
Total Images: 1,000 (subset of original 10,200)
Classes: 10 aircraft variants
Image Size: 224×224 pixels
Split:
- Training: 332 images
- Validation: 333 images
- Test: 335 images

Selected Aircraft Classes

The model classifies the following 10 aircraft variants:

707-320 - Boeing 707-320
737-400 - Boeing 737-400
767-300 - Boeing 767-300
DC-9-30 - McDonnell Douglas DC-9-30
DH-82 - de Havilland DH.82 Tiger Moth
Falcon_2000 - Dassault Falcon 2000
Il-76 - Ilyushin Il-76
MD-11 - McDonnell Douglas MD-11
Metroliner - Fairchild Metroliner
PA-28 - Piper PA-28

🏗️ Model Architecture

Transfer Learning Approach

Base Model: ResNet-18 (pre-trained on ImageNet)
Architecture: Deep Residual Network with 18 layers
Trainable Parameters: 11,181,642
Transfer Learning: Yes (fine-tuned backbone + new classifier)

Training Configuration

Batch Size: 32
Optimizer: Adam with differential learning rates
- Backbone Learning Rate: 0.0001
- Final Layer Learning Rate: 0.001
Loss Function: CrossEntropyLoss
Early Stopping: Patience of 10 epochs
Data Augmentation:
- Random horizontal flip
- Random rotation (±10°)
- Color jitter (brightness, contrast, saturation)
- ImageNet normalization

🚀 Installation

Prerequisites

Python 3.8 or higher
CUDA-compatible GPU (recommended)

Clone the Repository

git clone https://github.com/AhmedAl-Mahdi/Aircraft-Classifier.git
cd Aircraft-Classifier

Install Dependencies

pip install -r requirements.txt

💻 Usage

Training the Model

The complete training process is available in the Jupyter notebook:

jupyter notebook aircraft_classifier.ipynb

Using the Gradio Web Interface

Launch the interactive web interface:

python app.py

Then open your browser and navigate to the provided local URL to upload and classify aircraft images.

📈 Results

Performance Metrics

Test Accuracy: 87.17%
Weighted F1-Score: 0.8737
Training Accuracy: 100.00%
Best Validation Accuracy: 85.59%

Training Results

Epochs Trained: 17
Final Training Accuracy: 100.0000%
Final Validation Accuracy: 81.3814%
Training-Validation Gap: 18.6186%

Model Performance Analysis

The model demonstrates strong performance with:

High test accuracy of 87.17%
Good generalization despite some overfitting
Effective transfer learning from ImageNet pre-training
Successful fine-grained classification between similar aircraft types

Misclassification Analysis

Common misclassifications occur between visually similar aircraft types:

707-320 sometimes confused with DC-9-30 or MD-11
Similar commercial airliners show expected confusion patterns
The model maintains high confidence in correct predictions

🌐 Gradio Web Interface

The project includes a user-friendly web interface built with Gradio that allows users to:

Upload aircraft images via drag-and-drop or file selection
Get instant predictions with confidence scores
View the top predicted classes
Test the model with custom images

Features

Real-time Prediction: Instant classification results
Confidence Scores: Probability distribution across all classes
User-friendly Interface: Simple and intuitive design
Image Preprocessing: Automatic image resizing and normalization

📁 Project Structure

Aircraft-Classifier/
├── aircraft_classifier.ipynb    # Main training notebook
├── app.py                      # Gradio web interface
├── requirements.txt            # Python dependencies
├── README.md                   # Project documentation
├── LICENSE                     # MIT license
├── .gitignore                 # Git ignore rules
└── models/                    # Trained model files (created after training)
    └── aircraft_classifier.pth

🙏 Acknowledgments

FGVC-Aircraft Dataset: University of Oxford Visual Geometry Group for providing the comprehensive aircraft classification dataset
PyTorch Team: For the excellent deep learning framework that made this project possible
ResNet: He et al. for the residual network architecture that serves as the backbone of our model
ImageNet: For pre-trained weights enabling effective transfer learning

📚 References

Maji, S., et al. "Fine-Grained Visual Classification of Aircraft." arXiv preprint arXiv:1306.5151 (2013).
He, K., et al. "Deep Residual Learning for Image Recognition." CVPR 2016.
Deng, J., et al. "ImageNet: A Large-Scale Hierarchical Image Database." CVPR 2009.

📞 Contact

For questions or suggestions, please open an issue on GitHub or contact the repository owner.

Made with ❤️ for aircraft enthusiasts and machine learning practitioners