A Deep-Learning-Based Lightweight Model for Ship Localizations in SAR Images

Abstract

Ship detection and localizing its position are indispensable in maritime surveillance and monitoring. Until early 2000, ship detection relied on human intelligence, but with the fast-processing speed, artificial intelligence (AI), especially deep learning, has replaced manual intervention with automatic localization in tracking naval activities. Taking advantage of the continuous and cloud-free ocean observations of Synthetic Aperture Radar (SAR), recent studies have demonstrated some success in utilizing SAR data to localize ships using deep-learning and other AI methods despite the accuracy of the models being lower than the acceptable limit. However, the existing models are inherently complex and time consuming in addition to demanding an extensive computational resource, which pose a significant challenge when applied to satellite-based data. This study presents a computationally efficient deep-learning-based algorithm that has a wider applicability and improves the accuracy over the existing models for ship localization in SAR images. Training and testing of this algorithm were conducted using the SAR Ship Dataset, which contains ship chips with complex backgrounds extracted from Gaofen-3 and Sentinel-1 satellite data. It produced the localized ship’s position with bounding boxes in SAR images using the combined traditional computer vision and deep neural network configuration, which comprises a novel backbone network called Ship-Net or S-Net. The S-Net model has a thirteen-layer backbone feature extraction network and a four-layer regression network concatenated. Further, this study proposes a modified combined loss function for optimizing the model performance. A comparative analysis of the proposed S-Net model was done using the various pre-build model architectures and loss function combinations. The results showed that the S-Net model with a combined loss function yielded 94.88% precision and 79.68% recall, with 12.58% precision and 7.39% recall higher than the state-of-the-art Faster RCNN baseline model. The proposed S-Net model has a relatively higher performance than the existing state-of-the-art models for ship localization in SAR images and can become an efficient tool for ship localization in optical images with suitable architectural and training scheme modifications for better coastal surveillance and worldwide naval security.

Keywords

• Deep-learning
• maritime surveillance
• object detection
• ship detection
• SAR