基于超密集特征金字塔网络的SAR图像舰船检测

针对星载合成孔径雷达(synthetic aperture radar, SAR)图像舰船目标检测困难的问题,提出了一种基于超密集特征金字塔网络的检测算法。首先,利用残差神经网络提取原始图像特征,构建特征图。其次,跨尺度连接多个特征层获取超密集特征金字塔,建立多尺度的高层语义特征映射,增强特征传播和重用。然后，再利用区域建议网络提取每层金字塔的候选区域输入检测网络。最后,通过融合候选区域及其周边上下文信息,将检测网络注意力集中至海域以抑制虚警,并为分类器计算置信度和边框回归提供补充信息。多组仿真实验证明,所提网络框架设定合理且检测性能优越。

Ship detection in SAR images based on super dense feature pyramid networks

Aiming at the difficulty of ship target detection in space-borne synthetic aperture radar (SAR) images, a detection algorithm based on super dense feature pyramid networks is proposed. Firstly, residual neural network is used to extract features from original images and construct feature maps. Secondly, in order to enhance feature propagation and reuse, cross-scale feature layers are connected to obtain super dense feature pyramid and establish multi-scale high-level semantic feature mapping. Thirdly, candidate region is extracted from each layer of pdyramids by the regional proposal networks, and input into the detectim network. Finally, by fusing the candidate region and its surrounding contextual information to make the detector focus on the sea areas to suppress the false alarms, and provides supplementary information for the classifier to calculate confidence and bounding box regression. Simulation experiments show that the proposed network framework is reasonable and the detection performance is superior.