摘要
针对遥感图像下军事坦克检测模型体积大、计算量大等问题,提出一种轻量化的遥感军事坦克目标检测算法MSG-YOLOv7。首先,MSG-YOLOv7采用MobileNetv3作为主干网络,利用倒残差结构和自适应缩放的方法对特征进行提取,以减小模型的体积大小与运算量;其次,设计SD-MP结构来提高细节特征表达能力,解决因下采样操作导致的小目标特征丢失问题;最后,基于GCNet和深度可分离卷积设计出GD-ELAN模块,通过全局上下文建模来增强模型对长距离关系的感知,在轻量化的同时更有效地捕捉全局信息,提高模型的性能。实验结果表明,MSG-YOLOv7在公开的Google Earth遥感军事坦克数据集上的平均检测精度(AP)达到了99.02%,体积较原模型下降了60%,计算量为18.59 GFlops,FPS达到41,证明该模型适用于要求高性能、高速度和较小模型体积的遥感军事坦克检测场景中。
In view of the large volume and heavy computation of the military tank detection algorithm in remote sensing images,a lightweight remote sensing military tank target detection algorithm MSG-YOLOv7 is proposed.In the algorithm,the MobileNetv3 is taken as the backbone network,and the inverted residual structure and the adaptive scaling method are used to extract features,so that the volume and computation amount of the model are reduced.An SD-MP structure is designed to improve the ability of detailed feature representation,so as to eliminate the small target feature loss caused by downsampling operations.A module named GD-ELAN is devised based on GCNet and depthwise separable convolution.This module enhances the model's perception of long-distance relationships by global context modeling,capture global information effectively in a lightweight manner and improve the model performance.The experimental results show that the average precision(AP)of MSG-YOLOv7 of the proposed model on the public Google Earth remote sensing military tank dataset reaches 99.02%,with a volume reduction of 60%in comparison with that of the original,a computational complexity of 18.59 GFlops,and an FPS of 41,which proves that the model is applicable to remote sensing military tank detection scenarios that require high performance,high speed and small model volume.
作者
谢国波
吴陈锋
林志毅
XIE Guobo;WU Chenfeng;LIN Zhiyi(School of Computer Science and Technology,Guangdong University of Technology,Guangzhou 510000,China)
出处
《现代电子技术》
北大核心
2024年第19期47-54,共8页
Modern Electronics Technique
基金
国家自然科学基金资助项目(61802072)
南方电网委托课题(GDKJXM20230718)。