🛰️ Semantic Segmentation of Aerial Drone Imagery

This project focuses on semantic segmentation of aerial drone imagery using both CNN-Transformer hybrid models and pre-trained DeepLabV3 backbones.
The pipeline was designed to handle high-resolution drone images, achieving strong segmentation performance across multiple evaluation metrics.

---

📖 Table of Contents

• ✨ Key Highlights
• 🖼️ Example Outputs
• ⚙️ Methodology
• 📊 Results
• 📦 Tech Stack
• 🔮 Future Improvements
• 👤 Author

---

✨ Key Highlights

• 🧠 Computer Vision Model: Developed a semantic segmentation framework using a DeepLabV3 backbone integrated with a CNN-Transformer hybrid.
• 📈 Performance: Achieved 84% pixel-wise accuracy, 60% mIoU, and an F1-score of 74% on a dataset of 400 labeled aerial images.
• 🔄 Pipeline Engineering: Built a complete PyTorch pipeline for data preprocessing, training, and evaluation.
• 🛠️ Preprocessing: Implemented image tiling and mask remapping for efficient training on high-resolution drone imagery.
• 📊 Evaluation Metrics: Used mIoU, pixel accuracy, and F1-score for comprehensive model evaluation.

---

🖼️ Example Outputs

---

⚙️ Methodology

1. Data Preprocessing

• High-resolution aerial drone images split into tiles for GPU-efficient training.
• Remapped segmentation masks into consistent class labels.

2. Model Architectures

• DeepLabV3: Used pre-trained backbones for semantic segmentation.
• Hybrid CNN-Transformer: Designed a custom architecture combining convolutional layers for local feature extraction and Transformer blocks for global context.

3. Evaluation

• Metrics: Mean Intersection over Union (mIoU), Pixel Accuracy, F1-score.
• Evaluated on a public dataset of 400 labeled aerial drone images.

---

📊 Results

Model	Pixel Accuracy	mIoU	F1 Score
DeepLabV3 (pre-trained)	81%	55%	70%
CNN-Transformer (Hybrid)	84%	60%	74%

✅ Hybrid CNN-Transformer outperformed baseline DeepLabV3 across all metrics.

---

📦 Tech Stack

• Language: Python 3
• Frameworks/Libraries: PyTorch, Torchvision, NumPy, OpenCV, Matplotlib
• Deep Learning: DeepLabV3, Transformer layers
• Tools: Jupyter Notebook, CUDA-enabled GPU

---

🔮 Future Improvements

• ✅ Expand dataset with more diverse aerial imagery.
• ✅ Add support for multi-class segmentation beyond current labels.
• ✅ Deploy as a web-based visualization tool with FastAPI/Streamlit.
• ✅ Experiment with Vision Transformers (ViT) and Swin Transformers.

---

👤 Author

Jacob Almon

• LinkedIn
• GitHub

Svetya Koppisetty

• LinkedIn
• GitHub

Connor MacDonald

• LinkedIn
• GitHub

---