面向SAR图像的无锚框实时舰船目标检测算法

An Anchor Free Real Time Ship Target Detection Algorithm Based on SAR Image

锚框结构的舰船目标检测算法存在预设锚框与真实目标框难以精准匹配的问题,设计了一种基于合成孔径雷达(Synthetic Aperture Radar,SAR)图像的无锚框实时舰船目标检测算法。该算法以YOLOX-Nano(You Only Look Once X-Nano)框架为基础,在骨干网络单元嵌入改进Ghost模块和挤压激励(Squeeze and Excitation,SE)模块。路径聚合网络(Path Aggregation Network,PANet)与改进Ghost模块和自适应空间特征融合(Adaptively Spatial Feature Fusion,ASFF)模块集成后提高了模型的特征表达能力。以输入图像分辨率为320×320像素为基准,相较于单发多框检测器(Single Shot MultiBox Detector,SSD)和YOLOv3-tiny(You Only Look Once v3-tiny)模型,实验结果显示本文算法在合成孔径雷达舰船检测数据集(SAR Ship Detection Dataset,SSDD)上平均正确率达到94.5%,参数量为0.87×10^(6),浮点计算量为0.61×10^(9),能够实现高精度和低复杂度的SAR图像舰船目标检测。

Ship target detection algorithm with anchor has the problem that it is difficult to accurately match the preset anchor with the real target.An anchor free real-time ship target detection algorithm based on synthetic aperture radar(SAR)image is designed.Based on you only look once X-nano(YOLOX-Nano)framework,improved Ghost module and squeeze and excitation(SE)module are embedded in the backbone network unit.The feature expression ability of the model is improved after the integration of the path aggregation network(PANet)with the improved Ghost module and the adaptive spatial feature fusion(ASFF)module.The input image resolution is 320×320 pixels as the benchmark.Compared with the single shot multibox detector(SSD)and you only look once v3-tiny(YOLOv3-tiny)models,the experimental results show that the average precision of the algorithm in SAR ship detection dataset(SSDD)dataset is 94.5%,the parameter quantity is 0.87×10^(6),and the floating-point operations amount is 0.61×10^(9).Ship target detection in SAR images with high accuracy and low complexity can be realized.