RFP-based Faster R-CNN in Aerial Image Detection

Abstract

Aerial Image Detection is increasingly used in fields such as agriculture, natural disaster rescue, and topographic surveying. However, object detection in Aerial Image is still at a relatively low level. Generally speaking, there are the following challenges: 1) Lack of datasets; 2) A large number of targets in each image; 3) Object scales in the image are greatly different; 4) Each object category is uneven; 5) Image perspective is different (e.g., aerial view, elevation view and upward view). To improve the prediction effect of the network on such images, this paper proposes a deep learning framework for aerial static image detection based on improved Faster R-CNN—RFP-based Faster R-CNN. We add the Recursive Feature Pyramid (RFP) module to the Faster R-CNN for coping with the multi-scale small object in the image and effectively handle the occlusion. Bounded IoU Loss is used to replace the smooth LI Loss at Region of Interest (Rol) Pooling, and it avoids multiple boxes corresponding to the same Smooth LI Loss effectively. RFP-based Faster R-CNN achieves better performance. The proposed approach results show that the improved model achieves 20.7% mAP on the VisDrone2019-DET test set, which is higher than the most of classical object detection algorithms.