Abstract: Remote sensing images have the characteristics of complex background, large variations of object sizes and inter-class similarity, which lead to poor object detection results. An effective and robust remote sensing image object detection method based on Faster R-CNN is proposed. First, we introduce deformable convolution, feature modulation mechanisms and dilated convolution to construct a modulated feature adaptation network named MFANet, which can extract more accurate and complete object information. Second, a contextual feature pyramid network named CFPN is introduced to exploit richer and more discriminative feature representations. CFPN can solve the problems of insufficient high-level semantic information in the process of feature transfer and lack of effective communication between multi-size receptive fields. Finally, complete IoU (CIoU) loss is introduced into bounding box regression to further improve the accuracy of object detection. To verify the validity of the proposed method, we conduct experiments on public datasets DIOR, RSOD, and NWPU VHR-10. Experimental results show that compared with the Faster R-CNN with FPN method, IF-RCNN obtains an absolute gain of 8.43%, 7.5% and 8.0% in the average detection accuracy on the three datasets, respectively, which suggests that our proposed method is more effective and robust.

Key words: object detection, remote sensing image, feature pyramid, feature adaptation