摘要
目标建议作为目标检测任务中的预处理算法,可高效提取候选区域用于后续检测任务,提高检测效率。针对遥感图像飞机目标检测计算效率低下的问题,基于目标建议提出了一种飞机目标候选区域选取方法。首先基于多尺度局部非极大抑制算法,从多个尺度通过均值滤波提取局部梯度幅值极大的区域作为初始候选区域;然后利用图像边缘信息计算初始候选区域得分;最后根据飞机尺度特征设计尺度权重,结合非极大抑制剔除冗余窗口。利用机场遥感图像数据当提取1000个候选区域时取得了93. 7%的召回率,证明了该算法能够高效生成少量优质的候选区域,为进一步利用卷积神经网络等深度学习算法实现遥感图像飞机目标检测减少了计算量,提高了计算效率。
As preprocessing algorithm in target detection tasks,object proposal can efficiently extract a small number of candidate regions for subsequent detection tasks and improve detection efficiency. In view of the low computational efficiency of aircraft target detection in remote sensing images,a candidate region selection method for aircraft targets is proposed based on object proposal. Firstly,mean filter was used to extract the local gradient maximum area as the initial candidate regions from multiple scales based on the multi-scale local non-maximum suppression;secondly,the scores of coarse candidate regions were calculated with edge information of the images; finally,weights were designed according to the scale feature of the aircraft,and combined with non-maximum suppression to eliminate redundant windows. Using airport remote sensing image datasets,the recall rate achieves 93. 7% when 1000 candidate regions are extracted,it is proved that the algorithm can efficiently generate small quantity of good quality candidate regions,reduce the computational complexity and improve the computational efficiency for further utilizing deep learning algorithm such as convolution neural networks to detection aircraft targets in remote sensing images.
作者
唐小佩
杨小冈
刘云峰
李维鹏
TANG Xiao-pei;YANG Xiao-gang;LIU Yun-feng;Li Wei-peng(Rocket Force University of Engineering,Xi＇an Shanxi 710025,China;Rocket Force Military Deputy Office in Chengdu Area,Chengdu Siehuan 610036,China)
出处
《计算机仿真》
北大核心
2018年第11期45-50,共6页
Computer Simulation
基金
国家自然科学基金(61203189
61374054)
关键词
目标检测
目标建议
遥感图像
深度学习
Object detection
Object proposal
Remote sensing image
Deep learning
