Experimental evaluation of ship detection using U-Net with various backbone networks

Abstract

The growth in global trade has led to the growth in global ship traffic. Maritime security, safety and tracking has become more critical. Organizations and governments globally are developing applications to improve maritime surveillance. Among the widely used solutions, satellites are used to remotely scan and record images of waterways and shores. These image data feeds are processed using deep learning algorithms like Convolutional Neural Network (CNN) to detect ships with great precision. The challenges in this process are the image quality, size of ships, other varied noise within data and the computational requirements of handling and managing big data. The thesis work covers evaluating the U-Net model with popular transfer learning networks for semantic image segmentation on Airbus ship detection data. The backbones are pretrained on ImageNet dataset. Deep learning models were developed using Keras API with TensorFlow backend. The U-Net model with the selected backbones are compared with each other using standard metrics.