Object Detection in Aerial Remote Sensing Images with Multi-scale Feature Enhancement

Abstract

Abstract In target detection of satellite images, the significant differences in target scales can lead to many missed detections and false detections. To address this issue, an object detection algorithm based on improved YOLOv5s is proposed in this paper. Firstly, multiple dilated convolutions with different sampling rates are introduced in the Backbone network to aggrandize the capacity to extract detailed features of targets of different scales. Secondly, an adaptive feature fusion module is introduced based on the feature pyramid structure to fully utilize the characteristic information of different scales, and increase the detection capability of the network. Finally, experiments are carried out on the DIOR data set, and the proposed algorithm is demonstrated to be effective. Compared with traditional YOLOv5s, the proposed algorithm reduces missed detections and false detections and improves the overall accuracy of mAP (mean Average Precision) by 2.8%.