融合全局上下文注意力的遥感图像检测方法  

Remote sensing image detection method based onglobal context attention

作　　者：廖欢 朱文球[1] 雷源毅 徐轲 LIAO Huan;ZHU Wenqiu;LEI Yuanyi;XU Ke(Hunan University of Technology,Zhuzhou 412000,China)

机构地区：[1]湖南工业大学,湖南株洲412000

出　　处：《兵器装备工程学报》2024年第2期278-283,共6页Journal of Ordnance Equipment Engineering

基　　金：湖南省自然科学基金项目(2021JJ50058);湖南省教育厅开放平台创新基金资助项目(20K046)。

摘　　要：针对遥感图像场景复杂、目标尺寸不一、且小尺寸目标过多导致的检测精度不佳和出现漏检等问题,提出了一种融合全局上下文注意力的目标检测算法。该算法提出一种全局上下文注意力机制和YOLOv5中C3结构融合的模块,以提升网络捕捉图像全局特征的能力;通过Varifocal Loss损失函数来提升对密集、尺寸小的目标的检测性能;采用基于归一化的注意力模块,降低图像中不太显著的特征和权重,使网络能够达到更高的检测准确率;利用动态卷积学习各个维度的信息,让训练得到的模型在降低GFLOPs情况下,同时保持检测精度提升。在NWPU VHR-10数据集上实验结果mAP为96.0%、准确率为98.2%、召回率为94.9%,较原YOLOv5模型分别提升了1.8%、4.7%和2.2%,证明了所改进YOLOv5方法的有效性。A target detection algorithm that integrates global contextual attention is proposed to address the issues of poor detection accuracy and missed detections caused by complex remote sensing image scenes,varying target sizes,and excessively small target sizes.This algorithm proposes a module that integrates global context attention mechanism and C3 structure to enhance the network’s ability to capture global features of images;The Varifocal Loss function is used to improve the detection performance of dense small targets;A normalized attention module is adopted to reduce less significant features and weights in the image,enabling the network to achieve higher detection accuracy;By using dynamic convolution to learn information from various dimensions,the trained model can reduce GFLOPs while maintaining improved detection accuracy.The experimental results on the NWPU VHR-10 dataset showed mAP of 96.0%,accuracy of 98.2%,and recall of 94.9%,which were 1.8%,4.7%,and 2.2%higher than the original YOLOv5 model,respectively,demonstrating the effectiveness of the improved YOLOv5 method.

关 键 词：YOLOv5 遥感图像 Varifocal Loss 全局上下文注意力机制 动态卷积 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]