多尺度注意力特征增强融合的红外小目标检测新网络  

Multi-scale attention feature-enhanced fusion of a new network for infrared small object detection

作　　者：贾桂敏[1,2] 程羽 齐孟飞 JIA Guimin;CHENG Yu;QI Mengfei(Tianjin Key Lab for Advanced Signal Processing,Civil Aviation University of China,Tianjin 300300,China;College of Electronic Information and Automation,Civil Aviation University of China,Tianjin 300300,China)

机构地区：[1]中国民航大学天津市智能信号与图像处理重点实验室,天津300300 [2]中国民航大学电子信息与自动化学院,天津300300

出　　处：《中国安全科学学报》2024年第6期90-98,共9页China Safety Science Journal

基　　金：国家自然科学基金资助(61502498);天津市科研计划项目(2022KJ062);天津市自然基金资助(21JCYBJC00680);中央高校基本科研项目(3122019058)。

摘　　要：为提高红外成像中小目标检测的性能,提高低空空域监管能力,提出一种基于多尺度注意力特征增强融合的红外小目标检测新网络。首先,使用Resnet34提取红外图像的多尺度特征;其次,使用多尺度空间注意力特征增强模块(MFEM)来提高特征提取能力;然后,在逐级上采样过程中使用双通道注意力特征融合模块(DFFM),融合语义信息和细节信息,以更好地保护红外小目标的特征;最后,与其他方法对比,并以地/空红外弱小飞机目标视频序列检测为例测试真实场景。结果表明:新方法与现有方法相比,交互比(IoU)、F值和漏检率(FNR)的评分均获得改进;通过多尺度注意力特征增强融合可准确地定位到目标并生成精细的分割结果;MFEM能够同时利用多尺度上下文信息和空间注意力机制来突出红外小目标;DFFM通过给不同通道特征的集合赋予权重,得到最合适的特征图进行特征融合,从而提高检测性能。In order to improve the performance of small target detection in infrared imaging and the ability of low altitude airspace supervision,an infrared small target detection network based on multi-scale attention feature enhancement fusion was proposed.Firstly,Resnet34 was used to extract the multi-scale features of infrared images.Secondly,the multi-scale spatial attention feature enhancement module(MFEM)was used to improve the ability of feature extraction.Then,in the step-by-step up sampling process,the dual channel attention feature fusion module(DFFM)was used to fuse the semantic information and detail information to better protect the characteristics of infrared small targets.Finally,taking the video sequence detection of ground/air infrared dim small aircraft target as an example,the real scene test was carried out by comparing with other methods.The results show that compared with existing methods,the proposed method improves the scores of intersection over union(IoU),F-measure and false negative rate(FNR),and can accurately locate the target and generate good segmentation results.The DFFM can simultaneously use multi-scale context information and spatial attention mechanism to highlight infiared small targets.The DFFM assigns weights to sets of different channel features,thereby obtaining the most appropriate feature map for feature fusion and improving the detection performance.

关 键 词：红外图像 小目标检测 特征增强 特征融合 注意力机制 

分 类 号：X949[环境科学与工程—安全科学]