Ship_Detection

🚢 Ship Detection (Final Milestone)

VITMAV45 – Deep Learning in Practice with Python and LUA

Major Term Project – Airbus Ship Detection Challenge (Kaggle)

Megajánlott jegyért!

---

📘 Project Overview

This repository contains our final solution for the Deep Learning in Practice course. Our project tackles the Airbus Ship Detection Challenge on Kaggle, focusing on detecting ships in satellite images using a robust Two-Stage Deep Learning Pipeline.

In addition to the research notebooks, we have deployed the model as a Full-Stack Web Application allowing users to upload satellite images and receive real-time detection overlays.

---

🎯 Key Objectives & Achievements

• Milestone 1: Data exploration, visualization, and preparation of train/val/test splits.
• Milestone 2: Initial U-Net training, data loading optimization, and baseline evaluation.
• Milestone 3 (Final):
  • Advanced Architecture: Implemented a Two-Stage Pipeline (Classifier + Segmenter) to handle class imbalance.
  • Optimization: Reduced False Positives significantly by filtering empty images.
  • Performance: Achieved high F2 Score and IoU on the unseen test set.
  • Web Deployment: Containerized the solution using Docker, FastAPI, and React.

---

👥 Team Information

Team Name: 99_Problems_but_a_Batch_Aint_One

Name	Neptun Code	Role
Bologa Eduard	DAM4AV	Documentation, GitHub integration, pipeline architecture design, final report
Kozma Szabolcs András	TKGQWN	Data loading optimization, loss function tuning (Tversky), metric implementation
Pünkösti Györk	VCV3N5	Model training (Classifier & Segmenter), inference pipeline, visualizations, Dockerization

---

🧠 The Solution: Two-Stage Pipeline

Due to the extreme class imbalance (most satellite images contain only sea), a single segmentation model often produces false positives (“ghost ships”) on waves or clouds. We solved this with a two-step approach:

1. Stage 1: The Classifier (Filter)

• Model: ResNet34 (Binary Classifier)
• Input: 384x384 resized image.
• Goal: Rapidly determine if an image contains any ship.
• Result: Filters out ~70% of empty images before they reach the heavier segmentation model.

2. Stage 2: The Segmenter (Pixel-Level Detection)

• Model: U-Net with EfficientNet-B4 Encoder.
• Input: 768x768 full-resolution image (only if Stage 1 detects a ship).
• Loss Function: Tversky Loss ($\alpha=0.3$, $\beta=0.7$) to prioritize Recall (F2 Score).
• Goal: Precise pixel-wise segmentation of ship boundaries.

---

🚀 Optimization & Results

We implemented several key optimizations to improve performance and training speed:

• Fast Data Loading: Pre-processed RLE strings into list objects before the Dataset loop to remove CPU bottlenecks.
• Threshold Tuning: Automated script to find the optimal probability threshold (e.g., 0.35 instead of 0.5) that maximizes the F2 Score.
• Tversky Loss: Switched from Dice Loss to Tversky Loss to penalize False Negatives more heavily (crucial for the F2 metric).

📊 Final Performance (Test Set)

Metric	Score	Description
F2 Score	0.8977	(Primary Metric) Harmonic mean of precision and recall, weighing recall higher.
Precision	81.58%	Percentage of predicted ships that are actually ships.
Recall	92.08%	Percentage of real ships that were successfully detected.

---

💻 How to Run the Web App

We have containerized the application using Docker for easy deployment.

Prerequisites

• Docker Desktop installed and running.
• Model Weights: Ensure you have the trained model files (best_classifier.pth and best_ship_segmenter.pth).

Quick Start

1. Clone the repository:

   git clone https://github.com/99ProblemsButABatchAint1/Ship_Detection.git
   cd Ship_Detection
   
   

2. Add Model Weights:
   • Note: Large model files are ignored by git.
   • Manually copy your .pth files into the backend/models/ directory.
3. Run with Docker Compose:

   docker compose up --build
   

   (Note: If you run into network errors on University WiFi/VPN, try disconnecting the VPN or setting the MTU in Docker settings to 1300).

4. Access the App:
   • Frontend (React): http://localhost:5173
   • Backend API Docs: http://localhost:8000/docs

---

📂 Repository Structure

├── LICENSE
├── notebooks
│   ├── milestone1.ipynb
│   ├── milestone2.ipynb
│   ├── milestone.ipynb
│   ├── notebook.ipynb
│   ├── ship_detection_work_notebook
│   ├── ship_detection_work_notebook.ipynb
│   └── unet_ship_detection.pth
├── project
│   ├── backend
│   │   ├── app
│   │   │   ├── deps.py
│   │   │   ├── main.py
│   │   │   ├── __pycache__
│   │   │   │   ├── deps.cpython-313.pyc
│   │   │   │   └── main.cpython-313.pyc
│   │   │   ├── routes
│   │   │   │   ├── detect.py
│   │   │   │   ├── health.py
│   │   │   │   └── __pycache__
│   │   │   │       ├── detect.cpython-313.pyc
│   │   │   │       └── health.cpython-313.pyc
│   │   │   └── services
│   │   │       ├── detector.py
│   │   │       ├── detector_stub.py
│   │   │       └── __pycache__
│   │   │           ├── detector.cpython-313.pyc
│   │   │           └── detector_stub.cpython-313.pyc
│   │   ├── Dockerfile
│   │   ├── models
│   │   │   ├── best_classifier.pth
│   │   │   └── best_ship_segmenter.pth
│   │   ├── __pycache__
│   │   └── requirements
│   ├── docker-compose.yml
│   ├── Dockerfile.dev
│   ├── eslint.config.js
│   ├── index.html
│   ├── package.json
│   ├── package-lock.json
│   ├── public
│   │   └── vite.svg
├── README.md
├── final_milestone_documentation.pdf
├── src
│   ├── best_classifier.pth
│   ├── best_ship_segmenter.pth
│   ├── dataset.py
│   ├── inference.py
│   ├── losses.py
│   ├── metrics.py
│   ├── model.py
│   ├── __pycache__
│   │   ├── dataset.cpython-313.pyc
│   │   ├── losses.cpython-313.pyc
│   │   ├── metrics.cpython-313.pyc
│   │   ├── model.cpython-313.pyc
│   │   ├── rle.cpython-313.pyc
│   │   └── transforms.cpython-313.pyc
│   ├── rle.py
│   ├── train.py
│   ├── transforms.py
│   └── tune_thresholds.py
├── unet_ship_detection.pth
└── weights
    ├── best_classifier.pth
    └── best_ship_segmenter.pth

This site is open source. Improve this page.