【摘要】 针对高分辨率遥感图像中物体排布密集、尺度变化较大等特性，提出一种目标检测算法R-YOLOv5。算法在YOLOv5模型基础上首先将跨阶段局部扩张结构作用于主干网络，采用一种加强的特征提取方式，通过整合空洞卷积和密集连接，来缓解模型对密集分布目标的漏检问题；其次，在主干网络的瓶颈部分结合Transformer模块来增强特征的表达，突出目标区域；最后，引入多尺度特征融合模块，解决多尺度特征融合时存在的不一致性问题，以提高模型的检测效果。在公开的遥感图像检测数据集DIOR的实验结果表明，R-YOLOv5算法平均精度均值(mAP)达到80.6%,具有良好的检测性能。更多还原

【Abstract】 Aiming at the characteristics of dense distribution and large scale variation of objects in high-resolution remote sensing images, an object detection algorithm R-YOLOv5 is proposed. On the basis of YOLOv5 model, the algorithm firstly introduces Cross Stage Partial Dilated Network in the backbone network, which adopts an enhanced feature extraction method to alleviate the problem of undetected dense distributed targets by integrating dilated convolution and dense connection. Secondly, in the bottleneck part of the backbone network, the Transformer module is combined to enhance the expression of features and highlight the target area. Finally, multi-scale feature fusion module is introduced to solve the inconsistency problem in multi-scale feature fusion to improve the detection effect of the model. The experimental results on public remote sensing image detection dataset DIOR show that the MAP of R-YOLOv5 reaches 80.6%, which has good detection performance.更多还原