基于时空域特征融合的红外弱小目标检测研究  

Infrared Dim Target Detection Based on Time-Space Domain Feature Fusion

作　　者：崔书玮 武文波[1] CUI Shuwei;WU Wenbo(Beijing Institute of Space Mechanics&Electricity,Beijing 100094,China)

机构地区：[1]北京空间机电研究所,北京100094

出　　处：《航天返回与遥感》2024年第5期79-88,共10页Spacecraft Recovery & Remote Sensing

基　　金：中国空间技术研究院人才基金项目(WYRC2021HQM)。

摘　　要：针对红外图像有效像素少且信噪比低,在空间域下目标与背景和噪声难以区分的问题,提出了一种基于时空域特征融合和改进YOLOv5目标检测网络的红外目标检测方法。该方法通过三维残差结构构建时空域特征融合模块,实现对红外弱目标时空域特征的高效提取,降低红外图像空间域噪声对目标检测的干扰;通过引入CA(Coordinate attention)注意力机制改进YOLOv5卷积神经网络,增强模型对微弱目标的敏感性,解决弱目标检测中目标相对于背景非常微弱的问题。实验结果表明,该方法与YOLOv5s网络相比,精确率增加2.2%,召回率增加2.1%,交并比阈值为0.5时的平均精度值增加3.5%,验证了时空域特征融合方法可以提高红外弱小运动目标的检测精度。Infrared target detection faces challenges such as limited effective pixels,low SNR,and difficulty in distinguishing targets from background and noise in the spatial domain.In response,we propose a infrared target detection method based on a spatiotemporal feature extraction module and an improved YOLOv5 object detection network.This method utilizes a three-dimensional residual structure to construct a space-time domain feature extraction module,enabling efficient extraction of space-time domain features of dim infrared targets and reducing interference from spatial domain noise in infrared image target detection.Additionally,we introduce the Coordinate Attention(CA)mechanism into the YOLOv5 convolutional neural network to address the challenge of detecting extremely weak targets relative to the background in weak target detection and improve the detection capability for weak targets.Experimental results demonstrate that compared to the YOLOv5s network,our proposed algorithm achieves a 2.2%increase in precision,a 2.1%improvement in recall,and a 3.5%increase in mean average precision at intersection over union 0.5.These results validate that the space-time domain feature fusion method can enhance the detection accuracy of weak infrared moving targets.

关 键 词：目标检测 深度学习 红外小目标 时空域特征融合 注意力机制 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置] V19[自动化与计算机技术—控制科学与工程]