MASSNet: Multiscale Attention for Single-Stage Ship Instance Segmentation

Abstract

Maritime surveillance is essential in understanding, predicting, and ensuring the security of events in the complex marine environment. In this context, we have used instance segmentation techniques, which provides an accurate and efficient method for segmenting objects (Ships) in maritime surveillance. However, prevalent two-stage algorithms have limitations, including complex models, extended training time, and high memory consumption, making them impractical for real-world application. To address these challenges, we present an efficient solution called Multiscale Attention for Single-Stage Ship Instance Segmentation, or MASSNet. MASSNet uses the power of attention mechanisms to enhance multiscale feature extraction across various dimensions, resulting in a more refined and contextually-aware representation. This approach significantly improves segmentation accuracy and overall performance. In our extensive experiments, we evaluate the effectiveness of MASSNet on three challenging datasets: MariboatS, ShipInsSeg, and ShipSG. Our proposed model achieves mask Average Precision (mask AP) scores of 55.4%, 55.5%, and 74.1% on MariboatS, ShipInsSeg, and ShipSG datasets and outperforms other models such as YOLACT, SOLO and SOLOv2 architectures. MASSNet offers a robust and efficient solution for Ship Instance Segmentation, making significant improvement in the capabilities of maritime surveillance.