基于改进YOLOv8的SAR图像飞机目标检测算法  

Aircraft Detection in SAR Images Based on Improved YOLOv8

在线阅读下载全文

作  者:邱琳琳 朱卫纲 李永刚 邱磊 李炫潮 QIU Linlin;ZHU Weigang;LI Yonggang;QIU Lei;LI Xuanchao(Space Engineering University,Beijing 101000 China)

机构地区:[1]航天工程大学,北京101000

出  处:《电光与控制》2025年第3期101-110,共10页Electronics Optics & Control

摘  要:合成孔径雷达(SAR)图像飞机目标检测面临背景复杂、飞机目标较为微弱且尺度较小、不同成像条件下目标差异较大、目标结构不连续等挑战,提出一种新的飞机目标检测算法ATDM,用于提高复杂背景下SAR图像飞机目标的检测精度。使用YOLOv8s作为基线模型并改进了损失函数,算法包含三个重要的模块,即卷积注意力模块(CBAM)、全维度特征提取(ODFE)模块和可变形全局特征融合(DGFF)模块。为了增强网络在复杂背景下对飞机目标特征的提取能力,在基线网络的Backbone插入CBAM,从空间和通道两个维度学习飞机目标的特征;ODFE利用卷积核空间四个维度的动态性,通过并行策略沿核空间的四个维度学习不同类型飞机目标的特征,提升复杂背景下对散射特性较弱的飞机目标及小目标的检测能力;DGFF自适应调整卷积核的形状以适应不同成像条件下得到的飞机目标,并进行全局信息特征融合。最后,改进边界框回归损失函数为动态非单调聚焦损失函数WIoU,采用动态非单调聚焦机制,使用离群度评估锚框质量,降低SAR图像中可能存在的错误标注产生的影响。为了评估所提ATDM的性能,在SADD和高分三号SAR飞机数据集上开展实验,在两个数据集上得到的平均准确率(AP)分别达到95.4%和98.2%;消融实验结果与分析验证了所提出的三个模块及损失函数的有效性。此外,在与其他目标检测算法的对比中,所提算法也得到了最高的平均准确率。The aircraft detection in Synthetic Aperture Radar(SAR)images encounters several challenges including complex backgrounds dimand small-scaleaircraft targets big differences in targets under different imaging conditions and fragmented target structures.To solve the problems a novel aircraft target detection algorithm named Aircraft Target Detection Model(ATDM)for SAR images is proposed to improve the detection accuracy of aircraft targets in SAR images in complex backgrounds.Taking YOLOv8s as the baseline model the algorithm includes three key modules namely the Convolutional Block Attention Module(CBAM)Omni-Dimensional Feature Extraction(ODFE)module and Deformable Global Feature Fusion(DGFF)module along with an improved loss function.In order to improve the feature extraction ability of the network in complex backgrounds the CBAM is integrated into the backbone of the baseline network to capture aircraft target features across spatial and channel dimensions.The ODFE utilizes the dynamics of four dimensions of convolution kernel space namely the size of the kernel the number of input channels the number of output channels and the number of convolution kernels to extract features from different types of aircraft targets across the four dimensions by using the parallel operation strategy thereby enhancing the detection of aircraft targets especially small targets with weak scattering characteristics in complex backgrounds.The DGFF dynamically adjusts the shapes and sizes of convolution kernels to accommodate variations in the imaging conditions thereby facilitating global information feature fusion.Finally the bounding box regression loss function is improved to be a dynamic non-monotonic focusing loss function WIoU.The dynamic non-monotonic focusing mechanism is adopted and the outlier degree is used to evaluate the quality of the anchor frame to mitigate mislabeling effects in SAR images.In order to assess the performance of the proposed ATDM the experiments are conducted on SADD and Gaofen-3 SAR aircraft datasets

关 键 词:合成孔径雷达 飞机检测 复杂背景 全维度特征提取 可变形全局特征融合 

分 类 号:TP751[自动化与计算机技术—检测技术与自动化装置]

 

参考文献:

正在载入数据...

 

二级参考文献:

正在载入数据...

 

耦合文献:

正在载入数据...

 

引证文献:

正在载入数据...

 

二级引证文献:

正在载入数据...

 

同被引文献:

正在载入数据...

 

相关期刊文献:

正在载入数据...

相关的主题
相关的作者对象
相关的机构对象