基于改进YOLOv5的 SAR 图像飞机目标细粒度识别

针对合成孔径雷达(synthetic aperture radar,SAR)图像飞机目标细粒度识别中的小目标和多尺度检测问题，提出了 一种基于YOLOv5的改进SAR图像飞机目标识别算法。该方法首先对网络进行重构，加入小目标检测层，改善小目标的漏 检问题，提高目标定位精度。其次，在颈部网络中引入极化自注意力机制(polarized self attention,PSA),并使用双边特征金 字塔结构(bir-directional feature pyramid network,BiFPN)进行多层特征带权融合，提高对飞机目标散射信息的关注度和滤除 干扰信息。最后，使用SIoU(SCYLLA intersection over union)作为网络损失函数提高网络收敛速度和检测精度。利用SAR- AIRcraft-1.0数据集进行了算法有效性试验研究，实验结果表明，算法有效提升了飞机目标的检测精度，精确率、召回率、平均 精度均值分别达到92.6%、84.1%、90.1%。

Fine-grained recognition of aircraft targets in SAR images based on improved YOLOv5

This study proposes a synthetic aperture radar(SAR)image aircraft target recognition algorithm based on improved YOLOv5 for the small arget and multi-scale detection problems inthe fine-grained detection of aircraft targets in synthetic aperture radar images.Firstly,a small target detection layer is added to solve the leakage detection problem of small targets and improve the target localization accuracy.Secondly,the polarized self-attention mechanism is introduced into the neck network,and the bi-directional feature pyramid network(BiFPN)is used for multilayer feature band-weight fusion to improve the attention to the scattering information of aircraft targets.Finally,SCYLLA intersection over union(SIoU)is used as the network loss function to improve the convergence speed.In this paper,the effectiveness of the algorithm is tested by using the SAR-AIRcraft 1.0 dataset,and the improved YOLOv5 achieves92.6%recall,84.1%precision,and 90.1%mAP@0.5.Experimental results show that the proposed algorithm hashigher detection accuracy than other mainstream one-stage algorithms