关键词: 细粒度检测, 遥感图像, 旋转检测, 舰船目标

Abstract: Remote sensing image ship target detection has been extensive used in engineering. However, there are currently few studies on fine-grained detection tasks and a lot of research on coarse-grained detection techniques. Fine-grained detection in remote sensing ship photos has two types of difficulties：the first is that the target is dispersed at any angle, and the second is that coarse and fine-grained characteristics of the target are intermingled. Due to these difficulties, current fine-grained detection techniques have poor detection accuracy. This research proposes an enhanced Oriented R-CNN network-based fine-grained identification approach for remote sensing ship image rotating targets in order to address the aforementioned issues. Specifically addressing the issue of ship targets that can be found in any angle in remote sensing images, this research presents a novel approach to produce candidate boxes for rotating targets based on Oriented R-CNN network, which may efficiently remove background redundant information and enhance model performance. In order to properly retain the coarse and fine-grained characteristics of targets and increase the power of fine-grained detection, a feature refusion approach is presented in this study to incorporate FPN layers and multilevel features in the two-stage network feature mapping stage. The experimental results demonstrate that the modified model has a strong detection effect and a detection accuracy of 83.57% on HRSC2016, a publicly accessible fine-grained detection dataset for remote sensing ships.

Key words: fine-grained detection, remote sensing image, rotation detection, ship target