基于YOLOv5全局注意力和上下文增强的遥感图像目标检测方法

Detection Method of Remote Sensing Image TargetsBased on YOLOv5 with Global-aware and Context Enhancement

针对遥感图像目标密集排列,提出一种基于YOLOv5的全局注意力和上下文增强的遥感图像目标检测算法。首先,在主干网络CSPDarknet53的尾部将C3模块替换为Transformer Encoder模块,利用全局注意力捕获目标和周围环境信息;再引入上下文增强模块,利用不同的分支结构获得侧重于大、中、小目标的特征信息;后处理中将NMS替换为DIoU_NMS,有效去除密集排列目标的冗余框,改善检测效果。在RSOD数据集对网络进行测试,与原网络相比,精度提升了13.9%,在飞机数据集进行了消融试验,验证了模块的有效性。

On account of the dense arrangement of remote sensing image targets,a detection algorithm of remote sensing image targets is proposed on the basis of YOLOv5 with global-aware and Context Enhancement network.Firstly,the C3 module is replaced with the Transformer Encoder module at the tail of the backbone CSPDarknet53,and global attention is utilized to capture the targets and surrounding environment information;then,the Context Enhancement module is introduced to obtain the feature information focused on large,medium and small targets by employing different branching structures;after that,the NMS is replaced by DIoU_NMS in post-processing to effectively remove the redundant boxes of densely arranged targets and improve the detection effect.The network is tested on the RSOD dataset,and the precision is improved by 13.9%when compared with that of the original network.Ablation experiments are conducted on the aircraft dataset to verify the effectiveness of the module.