摘要
为提高航拍图像目标检测精度以及检测速度,提出了基于自适应阈值的改进CenterNet航拍图像目标检测算法。以目标的中心点作为关键点代替锚框进行分类和边界回归,设计自适应阈值预测分支对预处理结果进行筛选优化。同时设计了编码-解码结构的主干网络,通过可变形空洞卷积结构以及基于注意力机制的卷积连接结构,将浅层空间信息以及深层语义信息进行有效提取以及特征融合,提升了输出特征图质量。并通过结构化信息丢弃和利用误检、漏检目标构建新样本的方法实现数据增强,降低误检率及漏检率。在公开数据集NWPU VHR-10上进行实验,结果表明,与基于ResNet-50的CenterNet相比,本文算法的平均精度均值提升5.17%,交并比为0.50和0.75的平均精度分别提升了3.57%和3.61%,检测速度达45frame·s^(-1),取得了良好的检测精度和实时性的平衡。
In order to improve the accuracy and speed of aerial image target detection,an improved CenterNet aerial image target detection algorithm based on adaptive threshold is proposed.The center point of the target is used as the key point to replace the anchor box for classification and boundary regression,and an adaptive threshold prediction branch is designed to screen and optimize the preprocessing results.At the same time,the encodingdecoding network structure is designed.Through the deformable cavity convolution structure and the convolutional block attention-connection structure based on the attention mechanism,shallow spatial information,and deep semantic information are effectively extracted and fused.In addition,data enhancement is realized by discarding structured information and building new samples with false and missing detection targets,so as to reduce false and missing detection rates.Experiments are performed on the open data set NWPU VHR-10,the results show that compared with CenterNet based on ResNet-50,mean average precision of proposed algorithm increased by 5.17%,and intersection of union of 0.50 and 0.75 are improved by 3.57% and 3.61%,respectively.The detection speed reaches 45 frame·s^(-1),achieving good detection accuracy and real-time balance.
作者
许延雷
梁继然
董国军
陈壮
Xu Yanlei;Liang Jiran;Dong Guojun;Chen Zhuang(School of Microelectronics,Tianjin University,Tianjin 300072,China;Tianjin Key Laboratory of Imaging and Sensing Microelectrorde Technology,Tianjin 300072,China;Tianjin 712 Communication&Broadcasting Shareholding Co.,Ltd.,Tianjin 300457,China)
出处
《激光与光电子学进展》
CSCD
北大核心
2021年第20期184-193,共10页
Laser & Optoelectronics Progress
基金
天津市科技重大专项与工程计划项目(19ZXZNGX00060)。
关键词
图像处理
目标检测
卷积神经网络
自适应阈值
航拍图像
image processing
target detection
convolutional neural network
adaptive threshold
aerial image
