基于改进YOLOv5的遥感图像目标检测研究

Research on Object Detection in Remote Sensing Images Based on YOLOv5

针对遥感图像中背景复杂度高、目标尺寸多样所导致的目标检测精度低的问题,提出一种基于改进YOLOv5的遥感图像目标检测算法;采用ConvNeXt网络作为主干网络,结合了CNN的局部特性和Transformer的全局特性,克服了传统CNN在全局上下文信息处理和长距离依赖关系挖掘上的局限性,实现了对全局信息的有效捕获;引入SimAM注意力机制,在不增加网络参数的情况下推断出特征图的3D注意力权值,提高网络的稳定性以及抗干扰能力;采用CFP捕获全局长距离依赖关系以及遥感图像的局部关键区域信息,以及新颖的SIoU loss边界框定位损失函数和Soft-SIoU-NMS非极大值抑制方法,进一步提升了遥感图像实时检测的效果;在RSOD数据集上进行的测试,结果表明本算法相比于原网络提高了10.6%的平均精度,达到了94.2%。

For the problems of low detection accuracy in complex backgrounds and diversity of target sizes in remote sensing images,an improved YOLOv5-based algorithm for remote sensing image object detection is proposed,The ConvNeXt network is adopted as the backbone network,CNN's local features and Transformer s global features are combined to overcome the limitations of processing global contextual information and effectively capturing long-range dependencies for traditional CNNs,and achieve the effective capture of global information.The SimAM attention mechanism is introduced to deduce the 3D attention weights of the feature maps without increasing network parameters,enhancing the network stability and anti-interference ability Concurrently,centralized feature pyramid(CFP)is used to capture the global long-range dependencies,local key region information in remote sensing images,novel SIoU loss function,and Soft-SIoU-NMS non-maximum suppression metho,and further improve the real-time detection performance of sensing images.Testing on the RSOD dataset,the results show that compared to the original network,the average precision of the proposed method improves 10.6%,reaches,94.2%.