Supervised Adaptive-RPN Network for Object Detection in Remote Sensing Images

Object detection is one of the most important tasks in the field of very high resolution (VHR) remote sensing (RS) images understanding. Due to the characteristics of VHR RS images, the detection performance is always limited by class imbalance and intersection-over-union (IoU) distribution imbalance. To mitigate the adverse effects caused thereby, we propose a supervised adaptive-RPN (SA-RPN) model with the help of deep learning in this paper. First, we introduced a supervised multi-dimensional attention network to overcome the foreground-background class imbalance. It can help the network to highlight the foreground and suppress the background effectively. Second, we develop the adaptive-RPN to reduce the negative impact of the IoU distribution imbalance by adaptively selecting the size of the anchor. The positive experimental results on the public data set validate the usefulness of our SA-RPN model. Compared with the popular deep learning RS object detection methods, our method achieves improved performance.