采用空间注意力机制的红外弱小目标检测网络  

Infrared dim and small target detection network based on spatial attention mechanism

作　　者：鞠默然 罗海波[1,2,4,5] 刘广琦[1,2,3,4,5] 刘云鹏 JU Mo-ran;LUO Hai-bo;LIU Guang-qi;LIU Yun-peng(Shenyang Institute of Automation,Chinese Academy of Sciences,Shenyang 110016,China;Institutes for Robotics and Intelligent Manufacturing,Chinese Academy of Sciences,Shenyang 110016,China;University of Chinese Academy of Sciences,Beijing 100049,China;Key Laboratory of Opt-Electronic Information Processing,Chinese Academy of Sciences,Shenyang 110016,China;Key Laboratory of Image Understanding and Computer Vision,Shenyang 110016,China)

机构地区：[1]中国科学院沈阳自动化研究所,辽宁沈阳110016 [2]中国科学院机器人与智能制造创新研究院,辽宁沈阳110016 [3]中国科学院大学,北京100049 [4]中国科学院光电信息处理重点实验室,辽宁沈阳110016 [5]辽宁省图像理解与视觉计算重点实验室,辽宁沈阳110016

出　　处：《光学精密工程》2021年第4期843-853,共11页Optics and Precision Engineering

基　　金：中国科学院创新基金重点项目资助(No.Y8K4160401)。

摘　　要：红外弱小目标检测被广泛应用于预警、制导等国防领域中。然而,红外弱小目标所占像素少、缺少形状特征和纹理特征,使得红外弱小目标检测成为一个具有挑战性的课题。针对红外弱小目标检测,提出了一种简单高效的实时红外弱小目标检测网络。检测网络利用自适应感受野融合模块来增加小目标周围的上下文信息,并通过引入空间注意力机制来建立不同区域之间的相关性模型,使不同区域之间的相关性和紧凑性得到加强。为了提高检测网络对目标的定位和正负样本的判别能力,分别利用GIOU loss和Focal loss来设计损失函数。在3个红外弱小目标序列和单帧红外图像上进行实验,检测网络分别取得了91.62%,71.54%,81.77%和90.67%的AP值,且检测速度接近165 FPS。实验结果表明,该红外弱小目标检测网络对复杂背景和低信噪比条件下的红外弱小目标具有较好的检测效果。Infrared small target detection has been widely used for early warning,guidance,and in other fields of national defense.However,infrared small target occupies less pixels and lacks shape and texture features,which makes detection of infrared dim and small targets a challenging task.In this study,a simple and efficient real-time infrared small target detection network has been proposed.The detection network uses the adaptive receptive field fusion module to increase the contextual information around the targets.In addition,we used spatial attention mechanism to determine the relationship among different regions,which can strengthen the correlation and compactness among different regions.To improve the ability of network for locating the target and estimating the positive and negative samples,GIOU Loss and Focal Loss were used to design the loss function.The experiments were conducted on three infrared small target sequences and a single frame image set.The proposed network achieved 91.62%,71.54%,81.77%and 90.67%AP,respectively,and maintained high detection speed at approximately 165 FPS.The experimental results showed that the proposed infrared dim and small target detection network has good detection performance on the infrared small targets with complex background and low SNR.

关 键 词：红外弱小目标检测 注意力机制 卷积神经网络 深度学习 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术]