摘要
针对当前YOLOv3算法在遥感图像中复杂场景飞机漏检、误检等问题,提出一种基于自适应特征融合及多尺度输出的遥感图像飞机检测算法.该算法首先使用K-means++代替K-means算法对数据集进行聚类,解决了K-means初始聚类中心不稳定性;然后,在YOLOv3网络基础上增加了一个含有分辨率信息的尺度,更有利于检测小目标飞机;最后,在网络模型四尺度输出前增加了自适应特征融合层,解决了不同尺度的特征融合不充分以及减少或消除反向传导受到负样本的影响.实验结果表明,改进的YOLOv3算法在遥感图像上飞机检测精度达到96.17%,比YOLOv3算法精度提高了2.6%.
In view of the problems such as missing detection and false detection of complex scene aircraft in remote sensing image by the YOLOv3 algorithm,an aircraft detection algorithm based on adaptive feature fusion and multiscale output is proposed.Firstly,K-means++is used to cluster the data set instead of K-means algorithm,which solves the instability of the initial clustering center of k-means.Then,on the basis of YOLOv3 network,a scale with resolution information is added,which is more conducive to the detection of small target aircraft.Finally,an adaptive feature fusion layer is added before the four scale output of the network model,which solves the problem of insufficient feature fusion at different scales and reduces or eliminates the influence of negative samples on reverse conduction.The experimental results show that the detection accuracy of the improved algorithm reaches 96.17%in the remote sensing image,which is 2.6%higher than that of the algorithm.
作者
李众
白瑞君
洪军
李亚伦
王高
杨剑
LI Zhong;BAI Rui-jun;HONG Jun;LI Ya-lun;WANG Gao;YANG Jian(School of Software,North University of China,Taiyuan 030051,Shanxi China;Shanxi Information Industry Technology Research Institute Limited Company,Taiyuan 030051,Shanxi China;Collage of information and communication engineering,Taiyuan 030051,Shanxi China)
出处
《微电子学与计算机》
2021年第4期40-45,51,共7页
Microelectronics & Computer
基金
国家自然科学基金(61573323)
山西省自然科学基金(201901D111149)。
关键词
自适应特征融合
多尺度
遥感图像
飞机检测
adaptive feature fusion
Multiscale
remote sensing image
aircraft detection
