基于深度特征融合的红外弱小目标检测方法  

Infrared Dim Small Target Detection Method Based on Depth Feature Fusion

作　　者：马天凤 杨震[2] 罗勇[2] 庄超楠 MA Tianfeng;YANG Zhen;LUO Yong;ZHUANG Chaonan(School of Materials Science and Engineering,Henan Institute of Technology,Xinxiang 453000,China;School of Electrical and Information Engineering,Zhengzhou University,Zhengzhou 450001,China;Xuchang City Jian′an Power Supply Company,Henan Electric Power Company of State Grid,Xuchang 461143,China)

机构地区：[1]河南工学院材料科学与工程学院,河南新乡453000 [2]郑州大学电气与信息工程学院,河南郑州450001 [3]国网河南省电力供电公司许昌市建安供电公司,河南许昌461000

出　　处：《郑州大学学报（理学版）》2023年第3期65-72,共8页Journal of Zhengzhou University:Natural Science Edition

基　　金：国家自然科学基金项目(61903340);河南省重点研发与推广专项科技攻关项目(222102210158)。

摘　　要：红外弱小目标具有信噪比低、目标尺寸小、特征不明显等特点,加之场景复杂度不断提升,杂波干扰严重,导致现有的红外弱小目标检测方法在面对复杂场景时性能衰减。综合手工方法提取目标单一的显著特征及深度学习方法提取图像综合特征的优势,设计了基于深度学习的红外弱小目标深度特征融合检测网络模型。首先,模型利用多尺度自适应特征提取网络来提取红外图像中弱小目标的原始特征与平滑度图像中弱小目标的平滑度特征;其次,为提高目标显著度,提出了一种多层级联特征融合策略,实现特征提取网络中小目标原始特征与平滑度特征的融合;最后,利用多层级联特征融合映射网络对红外弱小目标进行特征映射与背景抑制,获得背景杂波被极大抑制的红外弱小目标特征映射图像。实验结果表明,同现有的基于深度学习与基于手工特征的检测方法相比,所提出的检测方法在各种复杂的场景中都拥有较高的准确率及较低的虚警率,同时拥有较快的检测速度。The detection of infrared dim and small target was difficcult because of the low signal-to-noise ratio,small target size,and insignificant features.In addition,the scene complexity and the clutter interference could also leads to the degradation of the performance of the existing infrared small target detection methods in complex scenes.Based on the advantages of the manual feature method to extract a single salient feature of the target and the deep learning method to extract a comprehensive feature of the image,a deep feature fusion detection network model of infrared dim and small target was designed.Firstly,the original features of small targets in infrared images and the smoothness features of small targets in smoothness images were extracted by a multi-scale adaptive feature extraction network.Secondly,in order to improve the saliency of the target,a multi-hierarchical feature fusion strategy was proposed to realize the fusion of the original feature and the smoothness feature of the small target in the feature extraction network.Finally,the multi-hierarchical feature fusion mapping network was used to perform feature mapping and background suppression of the infrared dim and small target,and the feature mapping image of the infrared dim small target with the background clutter greatly suppressed was obtained.Experimental results showed that,compared with the existing detection methods based on deep learning and manual features,the proposed detection method had higher accuracy and lower false alarm rate in various complex scenes,at a faster detection speed.

关 键 词：红外技术 弱小目标检测 深度学习 平滑度图像 特征融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]