SAR影像船舶目标检测技术研究

Research on SAR Ship Detection

针对SAR影像特征单一、小船舶目标检测召回率低和近岸目标虚警率高等问题,提出基于跨阶段局部聚合残差变换网络CSPResNeXt(Cross Stage Partial ResNeXt)的SAR影像船舶检测算法。该算法首先在浅层特征提取网络中加入了感受野模块RFB(Receptive Field Block)来模仿人类视觉感知,用以增强目标特征的可辨识性和网络对小尺寸船舶的适应性;其次采用多特征层双向加权融合和多尺度检测方法,对特征进行重组优化,提升目标检测能力;最后提出了相适应的损失函数的计算方法,并通过数据增强提升网络的鲁棒性。在HRSID和SSDD数据集上的实验结果表明,改进算法对密集小目标的检测效果得到了提升,有效降低了近岸强散射目标的误检率,检测速度和精度更加优越。

Aimed at the problems of single feature of ship targets,low recall rate of small ship and high false alarm rate of inshore targets in SAR images,a SAR ship detection algorithm was proposed based on Cross Stage Partial Aggregated Residual Transform Network(CSPResNeXt-50)in this paper.The receptive field block(RFB)is added to the shallow feature extraction network to simulate human visual perception,so that the identifiability of features and the adaptability of the network to small-scale ships can be enhanced.At the same time,the bi-directional weighted fusion of multi-feature layers and multi-scale detection methods are used to reorganize and optimize the features to improve the target detection ability.Finally,the calculation method of the corresponding loss function is proposed,and the robustness of the network is improved by data enhancement.Experimental results on the HRSID and SSDD datasets show that the proposed algorithm has improved the detection effect of dense-small targets,effectively reduced the false detection rate of strong near-shore scattering targets,and has superior detection speed and accuracy.