Int J Performability Eng ›› 2020, Vol. 16 ›› Issue (11): 1803-1813.doi: 10.23940/ijpe.20.11.p12.18031813

Previous Articles     Next Articles

Remote Sensing Object Detection via an Improved YOLO Network

Qinggang* Wu and Xueming Zhai   

  1. School of Computer and Communication Engineering, Zhengzhou University of Light Industry, Zhengzhou, 450001, China
  • Submitted on ; Revised on ; Accepted on
  • Contact: *E-mail address:

Abstract: It is a challenging problem to detect small and dense objects in remote sensing images. To address these problems, this paper presents a novel network based on YOLOv3 to detect objects in a one-stage framework, which contains two improvements. Firstly, K-means++ is utilized to determine the number of initial bounding boxes and aspect ratio dimensions, which is beneficial to adaptively adjust the parameters in YOLOv3 and to boost the small object detection in remote sensing images. Secondly, the Soft Non-Maximum Suppression (Soft-NMS) technique is introduced to reduce the confidence values rather than directly discard the candidate boxes whose Intersection-over-Union (IOU) values are larger than a predefined threshold. The Soft-NMS can increase the object detection accuracy for small and dense objects by suppressing the redundant boxes with lower confidence values. Extensive experimental results on the standard NWPU VHR-10 remote sensing dataset demonstrate that the proposed deep network can accurately and efficiently detect a variety of small objects compared with state-of-the-art object detection methods.

Key words: object detection, Soft-NMS, K-means ++, YOLOv3, remote sensing image