Ship Detection and Segmentation using Unet
Sagar Karki, Siddhivinayak Kulkarni
Abstract
Maritime surveillance is very crucial for every nation even if it is surrounded by land. The surveillance can not only inform about the threats and illegal activities but also manage the traffic and regulate a lot of movement on water. With more and more satellites with Synthetic-Aperture Radar (SAR) and other capabilities to capture high-resolution images being launched, automation of ship detection from these images is now a necessity. The experiments are performed on satellite image datasets where various encoders and loss functions for the Unet model are evaluated based on segmentation results over the test dataset. Our research and analysis of these encoders and loss functions for the Unet model lead us to certain conclusions as to which situation would benefit from which of the loss function and encoder and their combination. The research shows that the modes are robust and can perform well even with difficult backgrounds like clouds, waves etc. Our best model archived 0.823 on F2 score at different IOU thresholds on Airbus ship detection challenge test dataset.