在多分辨率场景下基于合成孔径雷达(SAR)图像进行多类典型目标识别,是SAR 图像信息解译的重要环节。基于YOLO-v4 网络模型,针对目前机载SAR 图像及目标信息的特点,提出一种应用于真实机载平台下多场景跨分辨率的实时检测处理架构。文中通过对多类目标进行双重检测,对样本数据量低的训练集进行数据增强,并将图像分割后的同类型目标信息进行合并,解决了多分辨率SAR 场景下目标尺度跨度较大的问题。实验结果表明:该方法能够在相关机载SAR数据集上达到六类目标(机场、桥梁、立交桥、汽车、装甲车、飞机)82. 8%的mAP 值,对后续机载SAR 复杂场景下更多类型目标的检测识别具有重要的借鉴意义。

In multi-resolution circumstances, multi-class typical target recognition based on synthetic aperture radar (SAR) pictures is an important aspect of SAR image information interpretation. Based on the YOLO-v4 network model, a real-time detection processing architecture is provided for multi-scene cross-resolution detection on real airborne platforms using the existing properties of airborne SAR images and target information. The problem of a vast span of target scales in multi-resolution SAR sceneries is addressed by double detection of several types of targets, data augmentation of training sets with low sample data volume, and combining the same type of target information after picture segmentation. The experimental results show that this method can achieve 82. 8% mean average precision (MAP) values on the relevant airborne SAR dataset for six types of targets (airports, bridges, overpasses, cars, armed vehicles and aircraft), which has important implications for the detection and recognition of more types of targets in subsequent airborne SAR complex scenes.

TN957