NPA2Net: A Nested Path Aggregation Attention Network for Oriented SAR Ship Detection
C Z Zhang, Peng Liu, Haipeng Wang, Ya‐Qiu Jin
Abstract
Ship detection in Synthetic Aperture Radar (SAR) images is crucial for various applications in both civilian and military domains. In recent years, there has been substantial progress in SAR ship detection, largely driven by advancements in deep learning-based methods. However, some issues still exist that need to be addressed. First, most current ship detection methods are anchor-based. These methods suffer from poor generalization and have a heavy computation load. Second, there is significant scale variation among different ships in SAR images. Most ships occupy only a few pixels in SAR images and are difficult to detect. Third, detecting inshore ships in SAR images is challenging, as it is easily susceptible to interference from harbors and buildings on the land. To address such issues, a novel anchor-free method named Nested Path Aggregation Attention Network (NPA2Net) is proposed for oriented SAR ship detection. Specifically, the anchor-free framework of Box Boundary-Aware Vectors (BBAVs) is applied to our method, which has higher detection efficiency. Moreover, the Nested Path Aggregation Module (NPAM) is proposed to fuse multiresolution features for detecting multiscale ships and a Ship Attention Module (SAM) is introduced to suppress the interference from complex backgrounds and highlight ships’ features. The experimental results conducted on three datasets for oriented SAR ship detection validate the effectiveness and generalization of NPA2Net.