基于深层次特征增强网络的SAR图像舰船检测

针对合成孔径雷达图像中舰船目标检测困难的问题，提出了一种基于深层次特征增强网络的多尺度目标检测框架.利用Darknet53提取原始图像特征，自上而下建立四尺度特征金字塔；特别设计基于注意力机制的特征融合结构，自下而上衔接相邻特征层，构建增强型特征金字塔；利用候选区域及其周边上下文信息为检测器计算分类置信度和目标分数提供更高质量的判定依据.所提算法在SSDD公开数据集和SAR-Ship自建数据集上的平均检测精度分别为94.43%和91.92%.实验结果表明，该算法设定合理且检测性能优越. 

Ship Detection in SAR Images Based on Deep Feature Enhancement Network

Aiming at the difficulty of ship target detection in synthetic aperture radar images, a multi-scale target detection framework based on deep feature enhancement network was proposed. Darknet53 was used to extract features from original images, and build a four-scale feature pyramid from top to bottom. A feature fusion structure based on attention mechanism was specially designed to connect adjacent feature layers from bottom to top, and rebuild enhanced feature pyramid. Then, the proposed method utilized the candidate region and its surrounding context information to provide a higher quality judgment basis for the detector to calculate the classification confidence and target score.The average detection precision of the proposed method on SSDD public data set and SAR-Ship self-built data set were 94.43% and 91.92% respectively. The experimental results show that the proposed network framework is reasonable and has superior detection performance.