Efficient Ship Detection in Synthetic Aperture Radar Images and Lateral Images using Deep Learning Techniques

Abstract

Over the past few decades, Synthetic Aperture Radar (SAR) imagery has become a primary means for high-resolution earth observation and a monitoring solution well suited for maritime surveillance. Ship detection (selecting the bounding boxes corresponding to ships) in SAR images plays a significant role in marine monitoring and in disaster relief. In the past, classical machine learning algorithms have been used towards this goal. Recently, in the field of object detection, the accuracy and detection speed have been significantly improved with the advent of deep learning (DL) techniques. However, such DL methods are less explored in the area of ship detection. In this paper, a salutary approach for improving ship detection in SAR images using advanced deep learning techniques is proposed. Various state-of-the-art benchmark models i.e. Faster-RCNN, YOLOv5, G-CNN and SSD are compared to assess their detection performance in various publicly available SAR datasets. Additionally, the best one among the trained models is used to detect lateral images of ships in real-time scenarios. The study is performed on a new custom-made Lateral Ship Detection Dataset (LSDD) developed in-house. Further, the best detection model is deployed in real-time tracking by integrating it with deepSORT tracking algorithm. The experimental results show the effectiveness of the DL models in ship detection and tracking applications in the wild.