Multi-Scale Feature Enhancement and Rotation-Aware Network for Ship Detection in Remote Sensing Images

• Zining Wu,
• Jing Zhang,
• Yongheng Tang,

Vol. 19, No. 12, December 31, 2025
10.3837/tiis.2025.12.007, Download Paper (Free):

• Ship Detection
• Remote Sensing Images
• Rotated objects
• Hybrid loss function

Abstract

To tackle the issues arising from intricate backgrounds in remote sensing images, varying ship target sizes, and the limitations of traditional horizontal bounding boxes in detecting inclined ship targets, this paper proposes a ship detection algorithm in remote sensing images based on multi-scale feature enhancement and rotation awareness. First, the convolutional block attention module (CBAM) is incorporated into the framework. This module adaptively captures inter-channel and spatial feature dependencies to emphasize critical information, suppresses interference from background information, and boosts the model's feature expression capability. Second, the atrous spatial pyramid pooling module (ASPP) aggregates multi-scale contextual features through dilated convolutions, integrating global contextual information to handle complex and diverse ship targets. Finally, the hybrid loss function is redesigned by incorporating ship rotation angles, strengthening the model's adaptability and robustness to rotational target variations. The synergistic effect of CBAM, ASPP, and the rotation-aware hybrid loss function improves feature expression, multi-scale perception, and rotation adaptability, achieving an effective balance between detection accuracy and efficiency. The detection accuracy on the HRSC2016 dataset reached 73.61%, exhibiting a competitive advantage over existing algorithms while effectively balancing the model's efficiency and performance.

Statistics

Cite this article