Hangar Detection from Satellite Images with Mask-RCNN and YOLOv2 Algorithms

Abstract

The contribution of this paper is twofold. It first proposes a new dataset of high resolution satellite images of hangars located at civil and military airports. It also presents a hangar detection problem results from satellite images using this new dataset obtained by Mask R-CNN and YOLOv2 algorithms. The satellite dataset contains one thousand pictures obtained from Google Earth at 8, 11, 17, 29 degrees angles from height of 500,800 and 1000 meters. Among 1000 hangar images 650 and 200 images are used for training and validation, respectively, while the remaining 150 images are used for detection test purposes. The total number of hangar object instants in the dataset images is about 3000. The detection of hangars is a challenging problem as the dataset contains camouflaged and non-camouflaged targets in different sizes. Among the two approaches used in the detection problem Mask R-CNN utilizes a regional based algorithm and enables instance segmentation with a bounding box. YOLOv2, on the other hand, is a regression based algorithm, used in real-time applications, and provides a bounding box only. The object detection accuracy in terms of Average Precisions by using Mask R-CNN and YOLOv2 algorithms to detect different sized camouflaged and non-camouflaged hangar objects was obtained as 72% and 74%, respectively.