Multiscale Ship Detection Method in SAR Images Based on Information Compensation and Feature Enhancement
Mingming Zhu, Guoping Hu, Hao Zhou, Shiqiang Wang
Abstract
With the development of synthetic aperture radar (SAR)-based imaging technology, SAR ship detection has achieved notable breakthroughs. Due to the scale diversity and class imbalance of ships, ship detection in SAR images is still a substantial challenge. To solve these problems, this paper proposes a multiscale ship detection method for SAR images based on information compensation and feature enhancement. To improve the feature representations of multiscale ships, an information compensation module (ICM) and a feature enhancement module (FEM) are embedded into a feature pyramid network (FPN). Specifically, the ICM is constructed to extract and aggregate diverse spatial contextual information. At the same time, the FEM is introduced to solve the feature-level imbalance of the FPN by integrating multilevel ship features. In addition, a gradient density parameter is introduced to solve class imbalance problems. Experiments on a high-resolution SAR image dataset (HRSID) show that the proposed method achieves a comprehensive detection performance, i.e., 60.6% average precision (AP) and 67.1 frames per second (FPS) and outperforms other state-of-the-art methods.