Multi-Scale Proposal Regions Fusion Network for Detection and 3D Localization of the Infected Trees

Abstract

Forest surveillance towers have the advantages of long observation time, wide observation range, stable and real-time observation. In this paper, a multi-scale proposal regions fusion network (MFRPN) is proposed for detecting the infected trees automatically on the enhanced images from the forest surveillance towers, which can solve the problem that small and large targets can’t be effectively detected on a single scale. The proposed MFRPN includes multi-scale images, three CNNs, three different RPNs, and proposal regions fusion model. In the proposed method, we train and run the scale-specific detectors in a multi-task fashion. And, to obtain the accurate spatial level location information of the infected trees, we achieve the three-dimensional (3D) coordinates localization of the digital elevation model (DEM) by using the principle of forest surveillance towers imaging and terrain elevation data. The experimental results show the detection accuracy achieves 91.63%, the detection time of a single image is 0.46 second, and the 3D localization error is less than 50m. The proposed network can realize the real-time detection and 3D localization of the infected trees.