摘要
无人机遥感系统在检测复杂背景下红外小目标时存在虚警过高的问题,结合卷积神经网络提出一种两阶段的无人机遥感系统红外小目标检测模型。第一阶段利用Unet神经网络学习红外图像中目标的深度语义特征与浅层位置特征,同时增强红外小目标信号,并抑制背景信号。第二阶段利用Faster R-CNN对第一阶段输出的图像进行分析,检测图中红外小目标的位置与边框。在公开的无人机遥感系统红外小目标检测数据集上完成了验证实验,结果表明该模型将三个复杂背景数据集下红外小目标的检测精度分别提高了13.2、9.8与13个百分点,每秒处理的帧数分别增加了11、14与13。
Aiming at the problem of high false alarm rate of infrared small target detection of unmanned aerial vehicle remote sensing system in complex environment,a two-stage infrared small target detection model of the unmanned aerial vehicle remote sensing system is proposed with combination of convolutional neural networks.In the first phase,the Unet neural network is taken advantage to learn the deep semantic features and shallow location features of targets in the infrared image,meanwhile,the infrared week and small target signal is enhanced and the background signal is suppressed.In the second phase,Faster R-CNN is utilized to analyze the output image of the first phase,to detect the location and bounding box of infrared small target.Validation experiment is carried on the public infrared small target detection dataset of the unmanned aerial vehicle remote sensing system,the results show that the detection precision of the proposed model for infrared small target increases by 13.2、9.8 and 13 percentage points on three complex background datasets respectively,and the processed frames per second increase 11、14 and 13.
作者
杨芳
王萌
YANG Fang;WANG Meng(Police scientific research department,Hebei Vocational College of Public Security Police,Shijiazhuang 050000,China;Department of Public Security Technology and tactics,Hebei Vocational College of Public Security Police,Shijiazhuang 050000,China)
出处
《光学技术》
CAS
CSCD
北大核心
2024年第1期120-128,共9页
Optical Technique
基金
河北省教育厅课题(ZC2022112)
河北省社会科学发展研究课题(20210201241)。
关键词
目标检测
遥感系统
无人机
深度神经网络
残差网络
弱小目标
红外热成像
remote sensing system
unmanned aerial vehicle
deep neural networks
residual networks
week andsmall target
infrared thermography
