Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (15): 76-86.DOI: 10.3778/j.issn.1002-8331.2301-0220

• Object Detection • Previous Articles     Next Articles

Improved YOLOv5 Object Detection Algorithm for Remote Sensing Images

YANG Chen, SHE Lu, YANG Lu, FENG Zixian   

  1. School of Geography and Planning, Ningxia University, Yinchuan 750021, China
  • Online:2023-08-01 Published:2023-08-01

改进YOLOv5的遥感影像目标检测算法

杨晨,佘璐,杨璐,冯自贤   

  1. 宁夏大学 地理科学与规划学院,银川 750021

Abstract: An improved YOLOv5 is proposed to address complex backgrounds and small objects missing detection in remote sensing images. Firstly, considering that the high-level feature map contains little small object information caused by down-sampling of convolutional neural networks, low-level feature is reused to increase the small target feature information. The EMFFN(efficient multi-scale feature fusion network) is used in the feature fusion stage instead of the original PANet(path aggregation network) to efficiently fuse the feature map information at different scales by adding jump connections and skip connections. Finally, a bidirectional feature attention mechanism(BFAM) including channels attention and pixel attention is designed to improve detection in complex background. To evaluate the proposed model, this paper uses two remote sensing image datasets, DIOR and RSOD. The experimental results show that the improved YOLOv5 model achieves 87.8% and 96.6% detection accuracy in the DIOR and RSOD datasets respectively, which is 5.2 and 1.6?percentage points better than the original YOLOv5 algorithm, effectively improving the detection accuracy of small targets in complex backgrounds.

Key words: remote sensing images, deep learning, object detection, YOLOv5, attention mechanism, multi-scale feature fusion, DenseDarkNet model

摘要: 针对遥感影像目标检测中复杂背景的干扰,小目标检测效果差等问题,提出一种改进YOLOv5(you only look once v5)的遥感影像目标检测模型。针对卷积神经网络下采样导致的特征图中包含的小目标信息较少或消失的问题,引入特征复用以增加特征图中的小目标特征信息;在特征融合阶段时使用EMFFN(efficient multi-scale feature fusion network)的特征融合网络代替原有的PANet(path aggregation network),通过添加跳跃连接以及跨层连接高效融合不同尺度的特征图信息;为了应对复杂背景带来的检测效果变差的问题,提出了一种包含通道与像素的双向特征注意力机制(bidirectional feature attention mechanism,BFAM),以提高模型在复杂背景下的检测效果。实验结果表明,改进后的YOLOv5模型在DIOR数据集与RSOD数据集中分别取得了87.8%和96.6%的检测精度,相较原算法分别提高5.2和1.6个百分点,有效提高了复杂背景下的小目标检测精度。

关键词: 遥感影像, 深度学习, 目标检测, YOLOv5, 注意力机制, 多尺度特征融合, DenseDarkNet模型