| 摘要: |
| 针对合成孔径雷达（Synthetic Aperture Radar,SAR）图像飞机目标检测算法存在模型复杂度较高、检测效果差、泛化能力弱等问题，提出了一种基于改进YOLOv8的SAR图像飞机目标检测算法。首先，针对SAR图像飞机目标较小的特点，剔除大目标检测层，重构特征提取网络和特征融合网络，降低模型计算量。其次，在主干网络引入可变形卷积（Deformable Convolutional Network，DCN），增强特征提取能力；在颈部网络引入全局注意力机制（Global Attention Mechanism，GAM）提高检测精度。最后，采用WIOU（Wise-IoU）损失函数提高收敛速度和回归精度。在SADD数据集（SAR Aircraft Detection Dataset）上实验结果显示，改进算法较原YOLOv8算法模型体积压缩59.66%，参数量降低61.18%，计算量减少18.29%，最高精度提高至98.1%。与其他算法相比，所提算法在保证较高检测精度的情况下大幅降低了模型体积、参数量和计算量，实现了模型复杂度和检测精度的平衡。 |
| 关键词: 合成孔径雷达 飞机目标检测 网络重构 可变形卷积 GAM注意力机制 |
| DOI：10.20079/j.issn.1001-893x.230515007 |
|
| 基金项目:国家级大学生创新创业训练计划（202011075013, 202111075012） |
|
| An Aircraft Target Detection Algorithm Based on Improved YOLOv8 in SAR Image |
| CHEN Yifang,c,ZHANG Shang,RAN Xiukang |
| (a.College of Electrical Engineering and New Energy;b.Hubei Province Engineering Technology Research Center for Construction Quality Testing Equipment;c.College of Computer and Information,China Three Gorges University,Yichang 443002,China) |
| Abstract: |
| To solve the problems of high model complexity,poor detection performance,and poor ability of generalization in synthetic aperture radar （SAR） aircraft target detection,a novel algorithm based on improved YOLOv8 is proposed.First,considering the small size of aircraft targets in synthetic aperture radar (SAR) images,the algorithm removes the large object detection layers and reconstructs the feature extraction network and feature fusion network to reduce model computation.Secondly, it introduces deformable convolutions into the backbone network to enhance feature extraction capability and incorporates the global attention mechanism(GAM) into the neck network to improve detection accuracy.Finally,it adopts the Wise-IoU (WIOU) loss function to enhance convergence speed and regression accuracy.Experimental results on the SAR Aircraft Detection(SADD) dataset demonstrate that the proposed algorithm achieves a 59.66% reduction in model size, a 61.18% decrease in parameter quantity,and an 18.29% reduction in computation,while achieving a maximum accuracy of 98.1%.By comparing the proposed algorithm with other approaches,the results indicate a significant reduction in model size,parameter quantity,and computation while maintaining high detection accuracy, validate the balanced trade-off between model complexity and detection accuracy. |
| Key words: SAR network reconstruction deformable convolutional networks global attention mechanism |
