基于改进YOLOv7的SAR图像小目标检测方法
Improved Small Target Detection Method for SAR Image Based on YOLOv7
杨可 1司占军 1张滢雪 1石金玉1
作者信息
- 1. 天津科技大学 人工智能学院,天津 300457
- 折叠
摘要
针对当前合成孔径雷达(Synthetic Aperture Radar,SAR)图像目标检测方法不能够适应不同尺寸目标、图像背景复杂导致的检测精度低等问题,本研究提出了一种基于改进YOLOv7的SAR图像小目标检测方法.该方法对特征提取网络进行改进,在骨干网络中使用SAC卷积帮助模型捕获不同尺度的目标信息,提高对小目标的特征提取能力.同时基于注意力机制,在原始目标检测头上嵌入DyHead模块,在降低复杂背景影响的同时更好地筛选小目标.此外,引入NWD损失函数与CIoU损失函数相结合,相比于原CIoU损失函数,NWD损失函数更加注重对小目标的处理,从而进一步提高了对小目标的检测能力.通过在数据集HRSID上的实验结果表明,改进后方法的mAP@0.5、mAP@0.95分别为93.5%、71.5%,相比基线模型分别提升了7.2%、7.6%,其能够有效地完成SAR图像小目标检测任务.
Abstract
In order to solve the problems that the current synthetic aperture radar(SAR)image target detection method cannot adapt to targets of different sizes,and the complex image background leads to low detection accuracy,an improved SAR image small target detection method based on YOLOv7 was proposed in this study.The proposed method improved the feature extraction network by using Switchable Around Convolution(SAConv)in the backbone network to help the model capture target information at different scales,thus improving the feature extraction ability for small targets.Based on the attention mechanism,the DyHead module was embedded in the target detection head to reduce the impact of complex background,and better focus on the small targets.In addition,the NWD loss function was introduced and combined with CIoU loss.Compared to the CIoU loss function typically used in YOLOv7,the NWD loss function pays more attention to the processing of small targets,so as to further improve the detection ability of small targets.The experimental results on the HRSID dataset indicate that the proposed method achieved mAP@0.5 and mAP@0.95 scores of 93.5%and 71.5%,respectively.Compared to the baseline model,this represents an increase of 7.2%and 7.6%,respectively.The proposed method can effectively complete the task of SAR image small target detection.
关键词
小目标检测/合成孔径雷达/YOLOv7/DyHead模块/可切换的空洞卷积Key words
Small target detection/Synthetic aperture radar/YOLOv7/DyHead module/Switchable Around Convolution引用本文复制引用
出版年
2024
NETL
NSTL
万方数据