针对常规模型在合成孔径雷达(synthetic aperture radar，SAR)图像中对近岸旋转舰船目标检测效果不佳的问题，提出一种基于全局特征融合的SAR图像旋转舰船目标检测方法。首先，通过全局注意力特征金字塔网络融合不同层级特征，缩短了底层特征向顶层特征的传递路径。其次，在图像块融合阶段加入位置编码，以减少降采样导致的定位信息损失。最后，采用旋转特征对齐网络生成高质量的锚点和旋转对齐特征，用于分类和坐标回归。所提方法在SAR舰船斜框检测数据集(rotated ship detection dataset in SAR images，RSDD-SAR)上旋转交并比为0。5时的平均检测精度达到了0。894 8，对近岸和离岸舰船都有着较好的检测性能。
Rotated ship target detection algorithm in SAR images based on global feature fusion
Conventional models are not effective in detecting inshore rotated ship targets in synthetic aperture radar(SAR)images.To solve this problem,a method of detecting rotated ship targets in SAR images based on global feature fusion is proposed.Firstly,the global attention feature pyramid network is used to fuse features of different levels,which shortens the transmission path from the bottom feature to the top feature.Secondly,positional embedding is added in the image block fusion stage to reduce the loss of location information caused by down-sampling.Finally,the rotated feature alignment network is used to generate high-quality anchor points and rotated alignment features for classification and coordinate regression.The proposed method achieves an average detection precision of 0.894 8 on the rotated ship detection dataset in SAR images(RSDD-SAR)dataset when the rotated intersection over union(IoU)is 0.5.The proposed method has good detection performance for both inshore and offshore ships.
万方数据
维普
