摘要
针对SSD算法检测遥感图像目标时存在容易漏检且检测精度低的问题,提出基于残差SSD网络的遥感图像目标检测算法。该算法在SSD网络结构的基础上,将基准网络模型VGG替换为残差网络模型ResNet-50,通过增加网络深度,充分提取遥感图像小目标数据集的底层特征,引入注意力模块,使感受野更关注目标特征,增强低层网络的信息表征能力,采用特征金字塔融合方法融合网络结构的高层语义特征和低层视觉特征,增强检测目标的定位能力。实验结果表明,该算法增强了复杂背景的干扰抑制性,提高了小目标的检测精度,比传统的SSD算法具有更强的检测性能。
In view of the fact that SSD(signal shot multibox detector)algorithm is prone to missing inspection and has low detection accuracy when detecting remote sensing image objects,a remote sensing image object detection algorithm based on residual SSD network is proposed.On the basis of the SSD network structure,the benchmark network model VGG is replaced with the residual network model ResNet-50.By increasing the network depth,the underlying features of the small object data set of remote sensing images are extracted sufficiently,and the attention module is introduced to make the receptive field pay more attention to the object features and enhance the information representation ability of the low-level network.The feature pyramid fusion method is used to integrate the high-level semantic features and low-level visual features of the network structure,so as to enhance the localization ability of the detection objects.Experimental results show that the proposed algorithm enhances the interference suppression of complex background,and improves the detection accuracy of small objects.Therefore,it has better detection performance than the traditional SSD algorithm.
作者
吕向东
彭超亮
陈治国
孙鹏飞
赵晓楠
徐旸
LÜXiangdong;PENG Chaoiang;CHEN Zhiguo;SUN Pengfei;ZHAO Xiaonan;XU Yang(Shandong Port Qingdao Port Group Co.,Ltd.,Qingdao 266001,China;CRRC Changjiang Transport Equipment Group Co.,Ltd.,Wuhan 430065,China)
出处
《现代电子技术》
北大核心
2024年第7期49-53,共5页
Modern Electronics Technique
关键词
SSD
残差网络
注意力模块
金字塔融合
遥感图像
小目标
高层语义特征
低层视觉特征
SSD
residual network
attention module
pyramid fusion
remote sensing image
small object
high-level semantic feature
low-level visual feature
