Is Satellite Image Target Camouflage Still Valid Under Deep Learning Target Detection?

Abstract

Satellite images can be used to observe a wild range of ground in a bird's eye view. In the past, researchers used hand-craft features to detect targets in satellite images. With the rapid growth of deep learning, neural networks such as Faster R-CNN, YOLO, SSD and RetinaNet can detect targets precisely and quickly. Traditional camouflage is designed to make important targets hard to identify. So whether the satellite image target camouflage is still effective with deep learning target detection? In this paper, we use YOLO v3 and RetinaNet to verify the effectiveness of camouflage and propose an improved YOLO v3 to enhance detection efficiency, which raise the detection speed from 34.5fps to 55.3fps in an image of 416*416. The experimental results show that the target camouflage in the satellite image has no effect under deep learning target detection methods. At the end of the paper, suggestions on how to improve the camouflage effect to resist deep learning detection are proposed.