基于改进YOLOv5-ResNet的海上舰船SAR图像快速检测
Rapid Detection of SAR Images of Naval Vessels Based on Improved YOLOv5-ResNet
龙昊 1张思佳 2周晶 3王冠4
作者信息
- 1. 海军大连舰艇学院 作战软件与仿真研究所,大连 116018;大连海洋大学 信息工程学院,大连 116018
- 2. 大连海洋大学 信息工程学院,大连 116018
- 3. 海军大连舰艇学院 作战软件与仿真研究所,大连 116018
- 4. 空军通信士官学校,大连 116600
- 折叠
摘要
在恶劣天气和海浪等自然因素的影响下,基于可见光数据进行舰船目标监测等手段往往难以有效开展,需要借助主动式微波成像卫星合成孔径雷达(SAR)进行图像解译.为了解决深度学习在处理数据集较小图像上无法准确提取特征及数据相似度较高的问题,基于YOLOv5-ResNet提出了一种跨尺度融合机制,重新定义损失函数.研究表明,识别SAR舰船目标的准确率有一定的提升:识别单目标舰船检测最高准确度达到93%,同比YOLOv5 提升4%,比YOLOv5-ResNet50 提升20%;在近岸舰船目标检测上,有效降低了由于数据集质量不佳、模型训练方法不当等造成误差率的非必要上升.
Abstract
Under the influence of natural factors such as bad weather and waves,it is often difficult to effectively carry out ship target monitoring based on visible light data and other means,which requires the use of active microwave imaging satellite synthetic-aperture radar(SAR)for image interpretation.To address the issue of inaccurate feature extraction by deep learning when dealing with small datasets and images,as well as the problem of high data similarity,a cross-scale fusion mechanism based on YOLOv5-ResNet is proposed to redefine the loss function.The research shows that there is a certain improvement in the accuracy of identifying SAR ship targets:the maximum accuracy of identifying single ships is 93%,which is 4%higher than YOLOv5 and 20%higher than YOLOv5-ResNet50.In near-shore ship target detection,it effectively reduces the unnecessary increase in error rate caused by poor data set quality and inappropriate model training methods.
关键词
合成孔径雷达图像/星载SAR图像/舰船目标检测/YOLOv5/ResNet/跨尺度融合Key words
SAR images/Space-borne SAR images/Ship target detection/YOLOv5/ResNet/Cross scale fusion引用本文复制引用
出版年
2024
NETL
NSTL
万方数据