Samsung Innovation Campus project

Marine Debris Detection using U-Net

Overview

This project focuses on detecting and segmenting marine debris from underwater sonar images using a U-Net model with a VGG16 backbone. The dataset includes various types of debris, and the model is trained to segment these objects accurately using semantic segmentation techniques.

Requirements

• Python 3.x
• TensorFlow
• Keras
• OpenCV
• Streamlit

Dataset

The dataset consists of 1868 sonar images and 1868 corresponding masks. The images are used for training a U-Net model, which segments different types of marine debris. The dataset includes 11 distinct classes of objects, such as:

• Wall
• Can
• Drink-carton
• Tire
• Valve
• Bottle
• Shampoo-bottle
• Propeller
• Hook
• Chain
• Standing-bottle

Model Details

• Architecture: U-Net with VGG16 backbone (pre-trained on ImageNet).
• Loss Function: Binary Cross-Entropy with Dice coefficient as an evaluation metric.
• Optimizer: Adam (learning rate 0.0001).
• Epochs: 40 epochs with early stopping.
• Evaluation Metrics: Mean Intersection over Union (IoU) and Dice Similarity coefficient.

Results

IoU (Training Set): 0.5351 IoU (Validation Set): 0.5353 Dice Similarity (Test Set): 0.8571

Model