基于改进YOLOv3的SAR舰船图像目标识别技术

Target Detection Technology for SAR Ship Images Based on Improved YOLOv3

合成孔径雷达(SAR)图像自动目标识别(ATR)技术是人工图像解译的关键技术之一。针对传统的SAR舰船目标检测算法大多受限于场景且泛化能力较差的问题,设计了一种基于改进YOLOv3网络的检测模型。将YOLOv3与DenseNet网络融合,使用稠密网络模块代替用于提取中小尺度特征的残差网络模块,通过训练得到模型的最优权重,实现端到端的目标检测。使用综合交并比(GIoU)损失代替交并比(IoU)边界框回归损失,提供更加准确的边界框位置信息,提高检测精度,采用中国科学院空天信息研究院制作的SAR图像船舶检测数据集进行测试。测试结果表明:与原YOLOv3算法相比,改进后的YOLOv3检测准确率提高了1.4%。

Automatic target recognition(ATR) of synthetic aperture radar(SAR) images is one of the key techniques for artificial image interpretation. However,traditional SAR ship target detection algorithms are mostly limited to the scene and poor generalization ability. In view of this,a detection model based on the improved YOLOv3 is designed. The YOLOv3 is fused with the DenseNet. The dense network block is used to replace the residual block so as to improve the feature extraction ability,and the optimal weight of the model is obtained through training so as to achieve end-to-end target detection. The GIoU loss is used to replace the IoU loss so as to provide more accurate boundary box location information and improve the detection accuracy. The detection dataset of SAR ship images produced by Aerospace Information Research Institute,Chinese Academy of Sciences are used for testing. The test results show that,compared with the original YOLOv3 algorithm,the detection accuracy of the improved YOLOv3 increases by 1.4%.