导航

ACTA AERONAUTICAET ASTRONAUTICA SINICA ›› 2023, Vol. 44 ›› Issue (5): 326738.doi: 10.7527/S1000-6893.2021.26738

• Electronics and Electrical Engineering and Control • Previous Articles     Next Articles

Object detection in UAV images based on multi-scale split attention

Guotao MAO1, Tianmin DENG1,2(), Nanjing YU3   

  1. 1.School of Traffic and Transportation,Chongqing Jiaotong University,Chongqing 400074,China
    2.School of Automation,Chongqing University,Chongqing  400044,China
    3.School of Shipping and Naval Architecture,Chongqing Jiaotong University,Chongqing 400074,China
  • Received:2021-12-03 Revised:2021-12-20 Accepted:2021-12-31 Online:2023-03-15 Published:2022-01-11
  • Contact: Tianmin DENG E-mail:dtianmin@cqjtu.edu.cn
  • Supported by:
    National Key Research & Development Program of China(SQ2020YFF0418521);Chongqing Science and Technology Development Foundation(cstc2020jscx-dxwtBX0019);Joint Key Research & Development Program of Sichuan and Chongqing(cstc2020jscx-cylhX0007)

Abstract:

With the development of Unmanned Aerial Vehicle (UAV) remote sensing technology, UAV aerial image object detection has become a core technology in the field of UAV applications such as traffic planning, military reconnaissance and environmental monitoring. To overcome the problem of difficulty in feature extraction due to many instances of small objects and complex background in UAV images, this paper proposes an object detection algorithm for UAV aerial images based on multi-scale split attention, i.e., MAS-YOLO. Firstly, the multi-scale split attention unit embedded in the bottleneck layer of the backbone net-work is used to establish the long-range dependency relationship between different scales of attention, so as to enhance the expression ability of key features and suppress the interference of background noise. Secondly, an adaptive weighted feature fusion method is designed, which dynamically optimizes the weight of each output feature layer and realize the deep fusion of shallow and deep features. Finally, experimental results on the VisDrone public data set show that the proposed method achieves 34.7% mean Average Precision (mAP), which is 2.8% higher than that of the baseline algorithm YOLOv5, and can also significantly improve the performance of UAV image object detection in complex background.

Key words: unmanned aerial vehicle image, computer vision, object detection, attention mechanism, adaptive weighted feature fusion

CLC Number: