The verification of complement for UAVs image feature with ground image in RiMaDIS

Abstract

Despite the development of deep learning, the number of application cases in the civil engineering field has not increased so much. For the reason, it is difficult to define the boundary conditions of the problem to be solved, and although there are various abnormalities to be detected. Additionally, the anomaly data may be less or not exist. For example, illegal dumping and illegal occupation in river patrols can take various forms depending on the context. Considering river patrols using UAV / AI now, we can assume a situation where there are few aerial images at the start stage. In this study, we verified whether the learning data could be complemented by learning together with the images obtained on the ground registered in the river maintenance database RiMaDIS, using Faster R-CNN. For ground image selection, images close to the feature space of aerial images were selected based on several criteria and methods. Among the proposed methods, ground images selected by the occupancy rate of the Bounding Box and the Deep Network (ShuffleNet, Inception v3) improved the average Precision.