Classification and Segmentation of Marine Related Remote Sensing Imagery Data Using Deep Learning

Abstract

Ship monitoring plays a crucial role in maritime safety, port administration, Ship traffic, maritime emergency and national defense. Using object detection methods U-Net and YOLOv2, image-based Ship detection has been put into practice but these methods have limitations, in U-Net at runtime, we can run the image only once which reduces its speed and in YOLOv2 all the anchor boxes are of same size, so objects with various sizes and shapes are difficult to detect. Hence to solve these issues a proposal with better techniques like YOLOv3 has been used to detect objects with super speed and various sizes of objects with the help of anchor boxes and U-net, which only requires a small number of training samples but offers high results for segmentation tasks due to its usage of a loss function for each pixel in the input image, this allows for simple identification of specific cells within the segmentation map. Hence the performance of these algorithms is measured to determine which of these two algorithms has more accuracy.