计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (20): 133-141.DOI: 10.3778/j.issn.1002-8331.2012-0064

• 模式识别与人工智能 • 上一篇    下一篇

面向遥感图像小目标检测的改进YOLOv3算法

王建军,魏江,梅少辉,王健   

  1. 1.西北工业大学 电子信息学院,西安 710129
    2.西北工业大学 第365研究所,西安 710129
  • 出版日期:2021-10-15 发布日期:2021-10-21

Improved YOLOv3 for Small Object Detection in Remote Sensing Images

WANG Jianjun, WEI Jiang, MEI Shaohui, WANG Jian   

  1. 1.School of Electronics and Information, Northwestern Polytechnical University, Xi’an 710129, China
    2.No.365 Institute, Northwestern Polytechnical University, Xi’an 710129, China
  • Online:2021-10-15 Published:2021-10-21

摘要:

针对YOLOv3目标检测算法在遥感图像小目标检测方面精度较低的缺点,提出了一种改进的YOLOv3目标检测算法——YOLOv3-CS。根据对backbone中不同尺度特征重要性的分析重构了backbone,即增加具有丰富位置信息的浅层特征对应的卷积层深度,以此增强backbone对小目标特征的提取能力,引入RFB结构增大浅层特征图的感受野来提升小目标检测精度,优化了anchor boxes及其分配原则。在RSOD数据集的实验结果表明,YOLOv3-CS算法与YOLOv3相比,mAP提高6.49%,F1提高4.85%,所需存储空间降低12.58%,其中backbone的改进和RFB的引入对小目标检测的精度提升较为明显,说明提出的目标检测算法在遥感图像小目标检测方面有较高的优势。

关键词: 小目标检测, YOLOv3, 遥感图像

Abstract:

Aiming at low accuracy of the YOLOv3 object detection model in small object detection, an improved YOLOv3 object detection model-YOLOv3-CS is proposed. Firstly, based on analyzing the importance of different scale features in the backbone network, the backbone network is reconstructed, that is to increase the depth of convolutional layer corresponding to the shallow feature with rich location information and improve the ability of small target feature extraction in the backbone network. Secondly, the RFB structure is introduced to increase the sensing field of shallow feature maps for improving the detection accuracy of small objects. Finally, the anchor boxes and their allocation principles are optimized. Experimental results on RSOD-dataset show that the YOLOv3-CS model can increase map by 6.49%, F1 by 4.85%, and model size by 12.58% in comparison with YOLOv3. Among them, the reconstruction of backbone and the introduction of RFB can improve the accuracy of small target detection obviously, which shows that the proposed target detection algorithm has higher advantages in the detection of small targets in remote sensing images.

Key words: small object detection, YOLOv3, remote sensing images