Aerial image change detection using dual regions of interest networks

Abstract

This paper presents a novel change detection network for identifying scene changes about pairs of aerial images. The aerial images are obtained with Unmanned Aerial Vehicles (UAVs), used to analyze “where and how” changes happened between pairs of images at different times. Even that aerial images could conveniently provide real time detailed information for land cover analysis, there still exist two challenging obstacles for coping with the change detection issue on aerial images. The one is that the paired aerial images captured at different times are roughly aligned due to the fact that the camera is mounted on a moving platform like UAVs. The other one is that season changes, light changes, and noise disturbance frequently happen, and they are useless or even impeditive in real applications. To conquer the problems, we propose a change detection networks named dual regions of interest networks to locate semantic change with object level, which could be easier coping with the above-mentioned compared with pixel-based methods. Moreover, we also introduce a pipeline to create a “Aerial change detection dataset” for the research of the change detection issues of aerial images analysis. Our evaluations on this benchmark dataset, “CDnet 2014 dataset”, and “AICD 2012 dataset” demonstrate the good detection and localization performance.