Inclined Image Recognition for Aerial Mapping using Deep Learning and Tree based Models
Muhammad Attamimi, Ronny Mardiyanto, Astria Nur Irfansyah
One of the important capabilities of an unmanned aerial vehicle (UAV) is aerial mapping. Aerial mapping is an image registration problem, i.e., the problem of transforming different sets of images into one coordinate system. In image registration, the quality of the output is strongly influenced by the quality of input (i.e., images captured by the UAV). Therefore, selecting the quality of input images becomes important and one of the challenging task in aerial mapping because the ground truth in the mapping process is not given before the UAV flies. Typically, UAV takes images in sequence irrespective of its flight orientation and roll angle. These may result in the acquisition of bad quality images, possibly compromising the quality of mapping results, and increasing the computational cost of a registration process. To address these issues, we need a recognition system that is able to recognize images that are not suitable for the registration process. In this paper, we define these unsuitable images as “inclined images,” i.e., images captured by UAV that are not perpendicular to the ground. Although we can calculate the inclination angle using a gyroscope attached to the UAV, our interest here is to recognize these inclined images without the use of additional sensors in order to mimic how humans perform this task visually. To realize that, we utilize a deep learning method with the combination of tree-based models to build an inclined image recognition system. We have validated the proposed system with the images captured by the UAV. We collected 192 images and labelled them with two different levels of classes (i.e., coarse- and fine-classification). We compared this with several models and the results showed that our proposed system yielded an improvement of accuracy rate up to 3%.
aerial mapping; image registration; deep learning; image classification; tree-based models;