Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (20): 167-175.DOI: 10.3778/j.issn.1002-8331.2303-0520

• Graphics and Image Processing • Previous Articles     Next Articles

Research on YOLOv7 Remote Sensing Small Target Detection Algorithm Integrating Attention Mechanism

YU Junyu, LIU Sunjun, XU Tao   

  1. School of Software Engineering, Chengdu University of Information Technology, Chengdu 610225, China
  • Online:2023-10-15 Published:2023-10-15

融合注意力机制的YOLOv7遥感小目标检测算法研究

余俊宇,刘孙俊,许桃   

  1. 成都信息工程大学 软件工程学院,成都 610225

Abstract: For remote sensing object detection, due to the fact that it mainly consists of densely distributed small targets, there are cases of missed and false detections during the detection process. Additionally, the detection process is also affected by significant differences in target scale and complex detection background. Therefore, this article proposes an improved YOLOv7 object detection method. First, by combining the idea of global semantic information and local semantic information, the centralized feature pyramid(CFP) is used to solve the problem of low detection efficiency in remote sensing images due to the dense distribution of targets and complex detection background. Secondly, in response to the uncertain distribution of small targets in remote sensing images and their insufficient feature representation ability, which can easily lead to missed and false detections during the detection process, this article introduces a mixed attention module ACmix to enhance the sensitivity of the network to small target detection, in order to improve the detection accuracy of small targets. Finally, the WIOU loss function is used to optimize the loss function in the original network to improve the network’s ability to locate the detected target. Experimental comparisons are conducted on publicly available remote sensing data, and the improved network shows an increase of 0.068, 0.061, and 0.098 in mAP for the three detection targets of aircraft, oil tanks, and playgrounds, respectively. The experimental results show that the improved YOLOv7 network performance is improved in the case of complex detection backgrounds and dense distribution of detection targets.

Key words: remote sensing image, target detection, small goals, loss function, YOLOv7

摘要: 针对遥感目标检测而言,因其主要是分布密集的小目标从而导致在检测过程中存在漏检误检的情况,其次在检测中还会受目标尺度差异显著和检测背景复杂带来的影响,因此提出一种改进YOLOv7的目标检测方法。通过结合全局语义信息与局部语义信息的思想,利用集中特征金字塔CFP(centralized feature pyramid)解决遥感图像因目标分布密集以及检测背景复杂导致检测效率较低的问题;针对遥感图像中的小目标分布不定并且其特征表现能力不足从而在检测过程中容易存在漏检、误检的现象,因此,通过引入混合注意力模块ACmix加强网络对于小目标检测的敏感度,以提升对小目标的检测精度;使用WIOU损失函数来优化原网络中的损失函数,提升网络对检测目标的定位能力。在公开的遥感数据中进行实验对比,改进后的网络对于三个检测目标飞机、油罐、操场的mAP分别提升了0.068、0.061、0.098,实验结果表明,在检测背景复杂,检测目标密集分布的情况下,改进的YOLOv7网络性能有所提升。

关键词: 遥感图像, 目标检测, 小目标, 损失函数, YOLOv7