摘要
红外小目标跟踪过程中由于背景、外界杂波等干扰,导致跟踪精确度和实时性欠佳,为此,提出基于深度学习的红外过采样扫描图像小目标跟踪算法。首先构建了红外过采样扫描图像模型,通过背景估计、形态学开运算,对图像中背景以及外界杂波进行多级滤除;然后增加设计特征融合模块和区域选取模块来改进孪生网络,生成融合特征图输入目标区域,通过分类和回归计算提高图像的特征表征能力和跟踪精度;最后建立损失函数训练孪生网络,输出红外过采样扫描图像小目标跟踪结果。实验结果表明,利用所提算法进行图像滤除后,信噪比能够高达35 dB,所提算法的区域重叠率较高、跟踪精度高,且算法的实时性强,帧率达到200 fps以上,整体跟踪效果好。
In the process of infrared small object tracking,the tracking accuracy and real-time performance are poor.Therefore,the algorithm of infrared over-sampling scanning image small object tracking based on deep learning is proposed.Firstly,the infrared oversampling scanning image model is constructed,used to filter the background and external clutter,then add the design feature fusion module and area selection module to improve the twin network,generate the fusion feature map input target area,and improve the feature representation ability and tracking accuracy through classification and regression calculation.Finally,the loss function is established to train the twin network and output the small target tracking results of infrared oversampled scanning images.The experimental results show that the proposed algorithm can be up to 35 dB,the proposed algorithm has high regional overlap rate,high tracking accuracy,strong real-time algorithm,the frame rate reaches more than 200 fps,and the overall tracking effect is good.
作者
姚迎乐
赵娟
Yao Yingle;Zhao Juan(Department of Information Engineering,Institute of Zhengzhou Industrial Application Technology,Zhengzhou 451150,China;Department of Mathematics and Computer Science,Wuhan Polytechnic University,Wuhan 430023,China)
出处
《国外电子测量技术》
北大核心
2023年第1期35-40,共6页
Foreign Electronic Measurement Technology
基金
河南省科技厅科技攻关支持项目(202102210361,182102210555)资助
关键词
多级滤波
改进孪生网络
特征融合
区域选取
红外小目标跟踪
multi-level filtering
improve the twin network
feature fusion
area selection
infrared small target tracking
作者简介
姚迎乐,硕士,副教授,主要研究方向为人工智能、深度学习、并行编译,E-mail:yaoyingle2022@163.com;赵娟,博士,讲师,主要研究方向为人工智能、信息隐藏。
