提出一种改进YOLOv5 网络,并将其用于SAR 图像目标识别。为了优化网络性能,文中进行了三个方面的改进:使用宽度比和高度比作为标注框之间的距离度量,并采用k-means 聚类方法生成先验锚点框,作为预测框优化时的框尺寸初始值;改进框回归损失函数,引入Scylla 交并比来代替竞争性交并比,以提高对密集分布目标的定位精度;改进置信度损失函数,使用焦点损失来替代二元交叉熵,以提高在复杂背景下的目标识别精度。基于MSAR 数据集,选择了YOLOv3、常规YOLOv5 作为对比网络,进行了大量的SAR 图像目标识别实验。实验结果表明,相比两种对比网络,改进YOLOv5 网络对各种目标均具有更高的识别准确率、召回率和F1 值,以及更高的综合指标平均精度值和平均精度均值。

An improved YOLOv5 network is proposed in this paper and is applied in SAR image target recognition. In order to optimize the performance of the network, three improvements are made as following. Firstly, width ratio and height ratio are used as the distance metric between labeled boxes, and k-means clustering method are used to generate a priori anchor box as the initial value of box size for prediction box optimization. Secondly, the regression loss function is improved in that CIoU is replaced by SIoU to improve the localization accuracy for densely distributed targets. Finally, the confidence loss function is improved in that binary cross entropy is replaced by Focal Loss to improve the target recognition accuracy in complex backgrounds. In this paper, based on the MSAR dataset, YOLOv3 and conventional YOLOv5 are selected as the comparison networks, and a large number of SAR image target recognition experiments are conducted. The experiment results show that the improved YOLOv5 network has higher recognition accuracy, recall rate, F1, AP and mAP for all types of targets compared with the two comparison networks.

TN957.51

国家自然科学基金资助项目(61890544,91748106)