A segmentation based approach for improved detection of ships from SAR images using deep learning models

Satellite image processing is one of the rapidly emerging fields in research today. Several objects such as buildings, roads, ships, boats etc. extracted from satellite images can be used further in applications involving tracking and surveillance. In sea and ocean images, ships can be found in different sizes which can complicate the task of finding a pattern or some regularity. It is comparatively easier to identify a ship in a homogenous environment rather than a heterogenous environment that consists of elements such as coasts, harbor, vessel, rocks, islands etc. This paper focuses on ship detection from Synthetic Aperture Radar images using deep learning models like CNN, ResNet, VGG and DenseNet. A segmentation approach that eliminates the noisy regions in the image is proposed in this paper. Performance of the different models with and without segmentation were evaluated based on F2-score. The ResNet model with segmentation has outperformed other models with an F2-score of 0.86.