Research on Aircraft Object Detection Based on Deep Learning

Abstract

At present, the hot topic in academia is how to extract information in remote sensing images quickly. According to the existing methods, to identify the aircraft targets in remote sensing images more effectively, the RSOD-Dataset annotated by Wuhan University was selected and constructed into PASCAL VOC format in this paper. Faster R-CNN and YOLOv3 were used to test the dataset respectively, and their underlying network structures are studied and the results are evaluated using four evaluation metrics. It was found that the metrics obtained with YOLOv3 were Pr=94.72%, Re=80.25%, AP=89.45%, and mAP=89.45% while the metrics obtained with Faster R-CNN were Pr=49.00%, Re=68.06%, AP=60.59%, and mAP=60.59%. For the dataset of this paper, YOLOv3 can be used to identify aircraft targets in remote sensing images more effectively.