Ship Detection Transformer in SAR Images Based on Key Scattering Points Feature Aggregation and Context Feature Refinement

• Published in: IEEE journal of selected topics in applied earth observations and remote sensing Vol. 18; pp. 17820 - 17836
• Main Authors: Yin, Yifei, Yang, Zhu, Shi, Hao, Meng, Fanyu, Li, Wei
• Format: Journal Article
• Language: English
• Published: IEEE 2025
• Subjects: Accuracy; Detection transformer; Detectors; Feature extraction; Marine vehicles; Radar polarimetry; Remote sensing; Scattering; scattering points; Semantics; ship detection; Synthetic aperture radar; synthetic aperture radar (SAR); Transformers
• ISSN: 1939-1404; 2151-1535
• DOI: 10.1109/JSTARS.2025.3580747

In recent years, deep learning algorithms have demonstrated significant advancements in the field of ship detection using synthetic aperture radar (SAR). Nevertheless, two primary challenges persist in the task of SAR ship detection: first, owing to the unique imaging mechanism, targets in SAR images are typically represented by scattering points, posing challenges for accurate feature extraction and leading to issues of inaccurate localization. Second, the detectors are susceptible to generating false alarms due to interference from the complex backgrounds of inshore scenes. In order to mitigate the issues mentioned above, a ship detection transformer based on key scattering points feature aggregation and context feature refinement is proposed. Specifically, considering that the ship targets exist in the form of scattering points in the SAR images, a key scattering points feature aggregation module is designed to mine and aggregate the key scattering points feature of ship targets. By this method, it is beneficial to generate more accurate feature representation for improving the localization performance of the detectors. Furthermore, to address the issue of excessive false alarms under complex background interference, a context feature refinement module is designed to augment the semantic representation and context information of feature maps. Extensive experiments are conducted on the two public datasets to substantiate the superiority of our proposed detector compared with other state-of-the-art methods.