摘要
单兵伪装目标与背景之间在颜色上有高度的相似性,目标具有高度复杂的姿态,而且存在遮挡问题,这些问题使得单兵伪装目标检测较传统目标检测有很大的挑战性。针对上述问题,提出基于偏振信息和RGB(Red,Green,Blue)信息的深度学习算法,同时构建单兵伪装目标偏振图像数据集CIP3K(Multicam型迷彩伪装数据集和Woodland型迷彩伪装数据集)。基于Faster R-CNN(Faster Region-Convolutional Neural Network)提出一种双流特征融合网络TSF-Net,其能够融合目标偏振特征信息和RGB特征信息。在CIP3K数据集上进行大量实验,用来测试TSF-Net模型与其他检测模型的性能。实验结果表明,相较于Faster R-CNN,TSF-Net模型在两个数据集上的平均检测精度分别提高了8.2个百分点和8.8个百分点,且优于一些主流目标检测模型。
There is a high degree of color similarity between the individual camouflage target and the background,the target has a highly complex posture,and there are occlusion problems,which make individual camouflage target detection more challenging than traditional target detection.In order to solve the above problems,a depth learning algorithm based on polarization information and RGB(Red,Green,Blue)information is proposed,and the polarization image dataset CIP3K(Multicam type camouflage dataset and Woodland type camouflage dataset)is constructed.Based on Faster R-CNN(Faster Region-Convolutional Neural Network),a dual-stream feature fusion network TSF-Net is proposed,which can integrate target polarization feature information and RGB feature information.A large number of experiments are carried out on the CIP3K dataset to test the performance of the TSF-Net model and other detection models.The experimental results show that,compared with Faster R-CNN,the average detection accuracy of the TSF-Net model on the two datasets is increased by 8.2 percentages and 8.8 percentages,respectively,and is better than some mainstream object detection models.
作者
王荣昌
王峰
任帅军
王勇
Wang Rongchang;Wang Feng;Ren Shuaijun;Wang Yong(Department of Information Engineering,PLA Army Artillery Air Defense Force College,Hefei 230031,Anhui,China;Key Laboratory of Polarized Light Imaging Detection Technology of Anhui Province,Hefei 230031,Anhui,China)
出处
《光学学报》
EI
CAS
CSCD
北大核心
2022年第9期185-197,共13页
Acta Optica Sinica
关键词
机器视觉
单兵伪装检测
偏振成像
卷积神经网络
数据集
machine vision
individual camouflage detection
polarization imaging
convolutional neural network
dataset
