基于改进YOLOv3的合成孔径雷达影像舰船目标检测

为了提高合成孔径雷达(synthetic aperture radar,SAR)影像舰船目标的召回率和准确率,降低漏检率,通过以YOLOv3(you olny look once)为检测框架,对锚点框(anchor boxes)生成机制进行改进,提出了利用K-median++生成anchors的聚类算法.结果表明不当的初始聚类中心会降低anchor boxes的平均交并比(mean intersection over union,meanIOU);同时由于SAR舰船数据集存在少量大尺寸box(离群数据点),因此在实验中使用中位数代替平均值,对簇群计算聚类中心,聚类后anchor boxes的meanIOU高达77.10％,在均值聚类算法(K-means clustering algorithm)基础上提高了3.7个百分点,并且减少了迭代次数,计算量得到大幅度降低.可见相比传统基于K-means的YOLOv3,检测效果有了明显提升,召回率达到92.21％,均值平均精度(mean average precision,mAP)达到93.56％,分别提高了2.55、3.82个百分点.

SAR Image Ship Target Detection Based on Improved YOLOv3

In order to improve the recall rate and accuracy rate of SAR image ship target and reduce the missed detection rate, a clustering algorithm using k-media ++ to generate anchors was proposed by taking YOLOv3 as the detection framework and improving the anchor box generation mechanism.The results show that the improper initial clustering center will reduce the MeaIOU of anchor boxes.At the same time, due to the existence of a small number of large-size boxes (outlier data points) in SAR ship data sets, median was used instead of average value in the experiment to calculate the clustering center of cluster groups. MeanIOU of anchor boxes after clustering was as high as 77.10%, which increased by 3.7 percentage points on the basis of k-means and reduced the number of iterations, and the calculation amount was greatly reduced.It can be seen that compared with the traditional k-means based YOLOv3, the detection effect has been significantly improved, with the recall rate reaching 92.21% and the mAP reaching 93.56%, respectively 2.55 and 3.82 percentage points higher.