摘要
在传统的空中目标检测算法中,需要进行人工设计特征来满足不同天空背景下的目标检测需求,对未来防空作战体系要求的自动化和智能化提出了挑战。为此本文在防空武器成像系统应用背景下,参照PASCAL VOC2007数据集格式建立了空中红外飞机数据集,将基于卷积神经网络的Faster R-CNN算法与R-FCN算法应用到空中红外飞机检测问题中,其次在Caffe框架平台下利用Faster R-CNN+VGG16/Res Net-101模型、R-FCN+Res Net-101模型,分别对自建数据集中测试集进行了检测。实验结果表明,在应对远距离弱小目标、云层遮挡、对比度低、目标截断等较难检测情况时,Faster R-CNN和R-FCN算法均能够有效地检测出空中红外飞机。本文良好的检测效果为解决空中红外飞机检测问题提供了更加简洁的思路。
In the traditional air target detection algorithm,artificial design features are needed to meet the target detec-tion requirements in different sky backgrounds,which poses a challenge to the automation and intelligence of future air defense combat system requirements.For this reason,in the context of the application of air defense weapon imaging sys-tem,the airborne infrared aircraft data set was established with reference to the PASCAL VOC2007data set format.The Faster R-CNN algorithm and R-FCN algorithm based on convolutional neural network were applied to the issue of airborne infrared aircraft detection.Then,under the Caffe framework platform,the test set of self-built dataset is tested by using the Faster R-CNN+VGG16/ResNet-101model and the R-FCN+ResNet-101model.The experimen-tal results show that both Faster R-CNN and R-FCN algorithms can effectively detect infrared aircraft in the air when it is difficult to detect the small and weak targets,cloud cover,low contrast and target truncation.The good detection effect of this article provides a more concise way to solve the issue of airborne infrared aircraft detection.
作者
姜晓伟
王春平
付强
JIANG Xiao-wei;WANG Chun-ping;FU Qiang(Electronic and Optical Department, Shijiazhuang Campus, Army Engineering University of PLA, Shijiazhuang 050003 , China)
出处
《激光与红外》
CAS
CSCD
北大核心
2018年第12期1541-1546,共6页
Laser & Infrared
关键词
卷积神经网络
深度学习
空中红外飞机
目标检测
convolutional neural network
deep learning
airborne infrared aircraft
target detection
