Detection of Military Targets from Satellite Images using Deep Convolutional Neural Networks

Abstract

Due to the varying size, orientation, and background of images in the defense sector, it is a daunting task to discern and distinguish the military targets in them. Multitudes of solutions have been proposed in this arena, yet there is a significant need for much better and flawless outputs. In this chapter, we expound on a two-level solution-Edge Boxes and Convolutional Neural Network (CNN) for the detection of targets in satellite imagery, Super resolution of the image using Dense-skip-connections. In the first level, the military objects are detected from the satellite image using Edge Boxes. In satellite imagery, the edge data of targets contains very prominent and concise attributes. The traditionally engineered features such as Histogram of Oriented Gradients, Hough transform and Gabor feature do not work well for huge datasets. However, the Edge Boxes technique generates contours around the target objects and discards the remaining. The output of this level is fed to the second level, wherein, the proposed targets undergo image super resolution. The presented deep learning model tends to inherently learn an end-to-end mapping between images of lower resolution and higher resolution. This level can be portrayed as one which takes a low-resolution input image and constructs an up-sampled high-resolution image as the output. Unlike traditional methods (sparse coding based method, bicubic method) that handle each component separately, this method aims to optimize all the layers at once. Furthermore, for assuaging the vanishing gradient problem that is common to very deep networks, Dense-skip-connections are employed. These enable the building of shorter paths directly within multiple layers. Though the proposed model has a light weighted structure, it exhibits state-of-the-art restoration quality.