Computer and Modernization ›› 2022, Vol. 0 ›› Issue (12): 88-94.

Previous Articles     Next Articles

Remote Sensing Image Object Detection Based on Improved MoCo

  

  1. (State Key Laboratory of Electronic Measurement Technology, North University of China, Taiyuan 030051, China)
  • Online:2023-01-04 Published:2023-01-04

Abstract: In the intelligent processing of satellite remote sensing images, there are some problems such as inconsistent standards and uneven data distribution, resulting in few effective samples and poor object detection effect. Aiming at this phenomenon, an object detection algorithm based on MoCo unsupervised contrast learning model is proposed. The framework of object detection adopts YOLOv5 with ResNet50 as the backbone network, and the weight of ResNet50 obtained by contrastive learning is used as a fixed value to participate in the detection task training of YOLOv5 downstream without gradient iteration. The contrastive learning experiment is carried out on AID Dataset, and the top-1 accuracy of the improved MoCo v2 is 95.888%. In the downstream detection task, using the TGRS-HRRSD Dataset, the accuracy of mAP@.5:.95 with the improved MoCo v2 pre-training weight is 67.8%, which is 5.6 percentage points higher than that without the pre-training weight. The results show that the improved MoCo comparative learning model is effective, and the detection accuracy is also improved in the downstream detection tasks after the comparative learning. 

Key words: unsupervised contrastive learning, remote sensing image detection, attention, YOLOv5