Deep Learning | Transfer Learning | YOLOv8 | FastAPI

An end-to-end computer vision system that classifies aerial objects as Bird or Drone and detects them in real-world images, videos, and live camera feeds. The project combines transfer learning for classification and YOLOv8 object detection with a FastAPI backend and lightweight web interface for interactive inference.

The system supports:

• Image classification
• Image object detection
• Video object detection
• Laptop webcam live detection
• Mobile camera real-time detection via browser

Binary classification using deep learning and transfer learning.

Models explored:

• Custom CNN baseline
• EfficientNetB0
• ResNet50V2
• MobileNetV2 (best performer)

📌 MobileNetV2 selected for deployment due to its speed and lightweight architecture.

The classifier outputs:

• Predicted class (Bird / Drone)
• Confidence score

Object detection is implemented using YOLOv8n, trained on a labeled dataset of aerial bird and drone images.

• 3,319 labeled images
• Bounding box annotations
• Bird and Drone classes

The detector identifies multiple objects in a frame and outputs:

• Bounding boxes
• Class labels
• Confidence scores

Example output:

Bird detected at (x1, y1, x2, y2) with 0.91 confidence
Drone detected at (x1, y1, x2, y2) with 0.87 confidence

The project includes a lightweight web interface powered by FastAPI and a simple HTML/CSS/JavaScript frontend.

The interface allows users to interact with the models without needing Python knowledge.

Users can upload an image and choose:

• Classification Mode
• YOLO Detection Mode

The system returns predictions along with bounding boxes if detection is selected.

Users can upload a video file.

The backend processes the video frame-by-frame using YOLOv8 and returns a fully annotated video with detections.

Live object detection can be performed using the system's webcam.

Each frame is processed by YOLOv8 and displayed in real time with bounding boxes.

The interface also supports mobile phone cameras via the browser.

Using the device camera:

• Frames are captured in the browser
• Sent to the FastAPI backend
• YOLOv8 performs inference
• Annotated results are streamed back

This allows real-time aerial object detection directly from a smartphone.

For classification:

• Resize images to 224 × 224
• Normalize pixel values
• Apply data augmentation

For object detection:

• Resize images to 640 × 640
• Convert annotations to YOLO format
• Perform augmentation

Classification Pipeline

1. Custom CNN baseline
2. Transfer learning models
3. Hyperparameter tuning
4. Evaluation with confusion matrix

Object Detection Pipeline

1. YOLOv8 training
2. Bounding box learning
3. mAP evaluation
4. Inference testing

For each request:

1. Input received from UI
2. Preprocessing applied
3. Model inference executed
4. Output returned

Outputs include:

• Predicted class
• Confidence score
• Bounding boxes
• Annotated images or videos

• TensorFlow / Keras
• Transfer Learning
• YOLOv8 (Ultralytics)

• OpenCV
• NumPy
• PIL

• FastAPI
• Python

• HTML
• CSS
• JavaScript

• Google Colab
• Jupyter Notebook

Detect drones and birds near airport runways to reduce bird-strike risks.

Monitor unauthorized drones in restricted airspace.

Track bird activity for ecological research.

Identify aerial objects in surveillance footage.

Assist in drone traffic monitoring systems.

Prediction: Bird
Confidence: 97.3%

Bounding boxes with labels:

Bird — 0.91
Drone — 0.87

MIT License

Contributions are welcome.

If you'd like to improve the system:

1. Fork the repository
2. Create a feature branch
3. Submit a pull request

Omprakash