基于YOLOv5-RTBBS的遥感卫星图像目标检测算法研究  

Research on Object Detection Algorithm of Remote Sensing Satellite Image Based on YOLOv5-RTBBS

作　　者：韩国东 范国玺[1] HAN Guodong;FAN Guoxi(Faculty of Engineering,Ocean University of China,Qingdao 266100,China)

机构地区：[1]中国海洋大学工程学院,山东青岛266100

出　　处：《电视技术》2025年第1期5-10,共6页Video Engineering

摘　　要：为解决遥感卫星图像目标检测中背景复杂、目标尺寸复杂和小目标检测困难等问题,提出YOLOv5-RTBBS检测算法,使模型在遥感图像分析中具有更高的精确度和健壮性。首先,将RepConv、Transformer Encoder和BiFPN模块整合到原始YOLOv5网络中,使不同尺度目标的检测精度得到提高。其次,通过引入BAM(Bottleneck Attention Module)注意力机制到C3模块来应对复杂背景区域的干扰。最后,将SIoU(Scaled-IoU)损失函数集成到YOLOv5s中以精确定位较小的目标。实验结果表明,改进后的YOLOv5s算法在遥感卫星图像数据集上的各项指标均有所提升,优于其他现有算法的性能。In order to solve the problems of complex background,complex target size and difficult detection of small targets in remote sensing satellite image detection,the YOLOv5-RTBBS detection algorithm is proposed in this paper to make the model more accurate and robust in remote sensing image analysis.Firstly,RepConv,Transformer Encoder and BiFPN modules are integrated into the original YOLOv5 network to improve the detection accuracy of different scale targets.Secondly,we introduce Bottleneck Attention Module(BAM)attention mechanism into C3 module to deal with the interference of complex background areas.Finally,the SIoU loss function is integrated into YOLOv5s to pinpoint smaller targets.The experimental results show that the improved YOLOv5s algorithm has improved the performance of various indicators on remote sensing satellite image dataset,which is better than other existing algorithms.

关 键 词：遥感卫星图像 注意力机制 目标检测 特征融合 深度学习(DL) 

分 类 号：TP311.1[自动化与计算机技术—计算机软件与理论]