📌 Bird vs Drone Classification & Object Detection

Deep Learning | Transfer Learning | YOLOv8 | Streamlit

This project builds a complete AI pipeline to classify aerial objects as Bird or Drone, and optionally detect them in real-world images using YOLOv8. The system supports image classification, object detection, and a deployable Streamlit UI.

⭐ 🚀 Project Features

✅ 1. Image Classification (Binary – Bird / Drone)

• Custom CNN baseline

• Transfer Learning:

  • EfficientNetB0
  • ResNet50V2
  • MobileNetV2 (best performer)

• Achieved up to 97% accuracy

• Includes confusion matrix, precision, recall, and F1-score

✅ 2. Object Detection (YOLOv8)

• YOLOv8n model trained on 3,319 labeled images

• Detects and labels multiple birds/drones in a single frame

• Outputs bounding boxes + class labels

• Achieved:

  • mAP50 ≈ 0.82
  • Precision ≈ 0.82–0.85
  • Recall ≈ 0.77–0.79

✅ 3. Streamlit Web App

• Upload an image
• Select Classification or YOLO Detection
• Get predicted class + confidence score
• See bounding boxes for YOLO detection

---

⭐ 📂 Project Structure

Bird_Vs_Drone/
│
├── notebooks/
│   ├── 01_EDA.ipynb
│   ├── 02_Custom_CNN.ipynb
│   ├── 03_Transfer_Learning.ipynb
│   ├── 04_YOLO_Training.ipynb
│
├── models/
│   ├── best_custom_cnn.h5
│   ├── best_mobilenet.keras
│   ├── best_effnetb0.keras
│   ├── best_resnet50v2.keras
│   ├── best.pt   # YOLO weights
│
├── app/
│   ├── app.py            # Streamlit App
│   ├── utils.py
│   ├── requirements.txt
│
├── object_detection_Dataset/
│   ├── train/images
│   ├── train/labels
│   ├── val/images
│   ├── val/labels
│   ├── test/images
│   ├── test/labels
│
├── classification_dataset/
│   ├── train/bird, drone
│   ├── val/bird, drone
│   ├── test/bird, drone
│
└── README.md

⭐ 📊 Model Performance Summary

🔹 Custom CNN

• Accuracy: 69%
• F1-score: 0.72
• 8–10 hidden layers

🔹 Transfer Learning Results

Model	Accuracy	Precision	Recall	F1-score
EfficientNetB0	0.972	0.968	0.968	0.968
ResNet50V2	0.972	0.968	0.968	0.962
MobileNetV2	0.972	0.968	0.968	0.968

⭐ MobileNetV2 selected as best model for deployment (fastest + lightest).

⭐ 🧠 YOLOv8 Detection Results

• mAP50: 0.82+
• mAP50-95: ~0.53
• Precision: 0.82–0.85
• Recall: 0.77–0.79

Outputs example:

Bird detected at (x,y) with 0.91 confidence  
Drone detected at (x,y) with 0.87 confidence

⭐ 🖥 Streamlit App Usage

🔧 1. Install Requirements

pip install -r requirements.txt

▶️ 2. Run App

streamlit run app.py

🧭 3. Features

• Upload an image

• Choose:

  • Classification Mode — MobileNet prediction
  • YOLO Detection Mode — Bounding box detection

• View results instantly

---

⭐ 📘 How It Works (High Level)

1. Preprocessing

• Resize to 224×224 for classification
• Resize to 640×640 for YOLO
• Normalize pixel values
• Data augmentation applied

2. Model Training

• Custom CNN baseline → moderate accuracy
• Transfer Learning → high accuracy
• YOLOv8n → bounding-box detection

3. Evaluation

• Confusion matrices
• Training curves
• mAP & precision-recall metrics
• Side-by-side model comparison

4. Deployment

• Streamlit web app
• Easy to upload images
• Real-time YOLO inference

---

⭐ 🎯 Real-World Applications

• 🛫 Airport bird-strike prevention
• 🕊 Wildlife monitoring
• 🎥 Security & defense surveillance
• 📡 Unmanned aerial vehicle detection
• 🔬 Environmental research

---

⭐ 📄 Technologies Used

• Python, TensorFlow, Keras
• Transfer Learning
• YOLOv8 (Ultralytics)
• OpenCV
• NumPy, Pandas, Matplotlib, Seaborn
• Streamlit
• Google Colab

---

⭐ 📜 License

MIT License

---

⭐ 🤝 Contributing

Pull requests are welcome. For major changes, please open an issue first.

---

⭐ 💬 Contact

Created by Omi Feel free to connect or share feedback!

---

If you want, I can also generate: ✅ A professional project thumbnail ✅ A requirements.txt ✅ A video script for your README Just say the word.