Abstract
Geo-spatial object detection in high-resolution satellite images has many applications in urban planning, military applications, maritime surveillance, environment control and management. Despite the success of convolutional neural networks in object detection tasks in natural images, the current deep learning models face challenges in geo-spatial object detection in satellite images due to complex background, arbitrary views and large variations in object sizes. In this paper, we propose a framework that tackles these problems in efficient and effective way. The framework consists of two stages. The first stage generates multi-scale object proposals and the second stage classifies each proposal into different classes. The first stage utilizes feature pyramid network to obtain multi-scale feature maps and then convert each level of the pyramid into an independent multi-scale proposal generator by appending multiple region proposal networks (RPNs). We define scale range for each RPN to capture different scales of the target. The multi-scale object proposals are provided as input to the detection sub-network. We evaluate proposed framework on publicly available benchmark dataset, and from the experiment results, we demonstrate that proposed framework outperformed other reference methods