Stepwise Locating Bidirectional Pyramid Network for Object Detection in Remote Sensing Imagery

Abstract

Recently, optical object detection has made significant advancements in the field of remote sensing. However, small-scale object detection is still a major challenge in optical remote sensing image interpretation. Therefore, this letter proposed a novel object detection method called stepwise locating bidirectional pyramid network (Sw-LBPN) to heighten the ability of remote sensing image object detection. Precisely, a stepwise locating attention scheme is proposed to highlight useful information and suppress useless ones of objects step by step at the feature channel level for large-scale remote sensing images. To effectively realize multiscale feature aggregation, a simplified bidirectional feature pyramid network (SBFPN) is designed. Moreover, the skip connection is leveraged in the middle level of SBFPN, aiming at offsetting and reusing small-scale object information. Several experiments on the measured object detection in optical remote sensing images (DIOR) and Northwestern Polytechnical University very high resolution 10-class remote sensing images (NWPU VHR-10) datasets demonstrate the effectiveness and the superiority of the proposed method compared with some state of the arts.