《计算机应用》唯一官方网站 ›› 2022, Vol. 42 ›› Issue (12): 3723-3732.DOI: 10.11772/j.issn.1001-9081.2021101802

• 人工智能 • 上一篇    

基于改进YOLOv3的遥感图像小目标检测

冯号1, 黄朝兵1(), 文元桥2   

  1. 1.武汉理工大学 信息工程学院,武汉 430070
    2.武汉理工大学 智能交通系统研究中心,武汉,430063
  • 收稿日期:2021-10-22 修回日期:2022-01-10 接受日期:2022-01-14 发布日期:2022-01-19 出版日期:2022-12-10
  • 通讯作者: 黄朝兵
  • 作者简介:冯号(1996—),男,重庆人,硕士研究生,主要研究方向:信息处理、图像处理与识别
    文元桥(1975—),男,湖北松滋人,教授,博士,主要研究方向:水上交通安全、智能船舶。
  • 基金资助:
    国家自然科学基金资助项目(52072287)

Remote sensing image small target detection based on improved YOLOv3

Hao FENG1, Chaobing HUANG1(), Yuanqiao WEN2   

  1. 1.School of Information Engineering,Wuhan University of Technology,Wuhan Hubei 430070,China
    2.Intelligent Transportation Systems Research Center,Wuhan University of Technology,Wuhan Hubei 430063,China
  • Received:2021-10-22 Revised:2022-01-10 Accepted:2022-01-14 Online:2022-01-19 Published:2022-12-10
  • Contact: Chaobing HUANG
  • About author:FENG Hao, born in 1996, M. S. candidate. His research interests include information processing, image processing and recognition.
    WEN Yuanqiao, born in 1975, Ph. D., professor. His research interests include water traffic safety, intelligent ships.
  • Supported by:
    National Natural Science Foundation of China(52072287)

摘要:

YOLOv3算法被广泛地应用于目标检测任务。虽然在YOLOv3基础上改进的一些算法取得了一定的成果,但是仍存在表征能力不足且检测精度不高的问题,尤其对小目标的检测还不能满足需求。针对上述问题,提出了一种改进YOLOv3的遥感图像小目标检测算法。首先,使用K均值聚类变换(K-means-T)算法优化锚框的大小,从而提升先验框和真实框之间的匹配度;其次,优化置信度损失函数,以解决难易样本分布不均衡的问题;最后,引入注意力机制来提高算法对细节信息的感知能力。在RSOD数据集上进行实验的结果显示,与原始的YOLOv3算法、YOLOv4算法相比,所提算法在小目标“飞机(aircraft)”类上的平均精确率(AP)分别提高了7.3个百分点和5.9个百分点。这表明所提算法能够有效检测遥感图像小目标,具有更高的准确率。

关键词: 小目标检测, YOLOv3, K均值聚类变换, 置信度损失函数, 注意力机制

Abstract:

YOLOv3 (You Only Look Once version 3) algorithm is widely used in target detection tasks. Although some improved algorithms based on YOLOv3 have achieved some results, there are still problems of insufficient representation ability and low detection accuracy, especially for the detection of small targets. In order to solve the above problems, a small target detection algorithm for remote sensing images based on YOLOv3 was proposed. Firstly, K-means Transformation (K-means-T) algorithm was used to optimize the size of anchor box, so that the matching degree between the priori box and ground truth box was improved. Secondly, the confidence loss function was optimized to solve the problem of uneven distribution of hard and easy samples. Finally, attention mechanism was introduced to improve the algorithm’s ability to perceive the detailed information. Results of the experiments carried out on RSOD dataset show that compared with the original YOLOv3 algorithm and YOLOv4 algorithm, the proposed algorithm has the detection Average Precision (AP) on the small target class “aircraft” increased by 7.3 percentage points and 5.9 percentage points respectively, illustrating that the proposed improved algorithm can detect small targets in remote sensing images effectively, with higher accuracy.

Key words: small target detection, YOLO (You Only Look Once)v3, K-means Transformation (K-means-T), confidence loss function, attention mechanism

中图分类号: