Exploiting Full-Scale Feature for Remote Sensing Object Detection Based on Refined Feature Mining and Adaptive Fusion

Abstract

Object detection for remote sensing images remains a challenging problem. In this paper, we proposed an effective remote sensing detection network (RS-Net) based on YOLOv4, which greatly solves the difficulties of remote sensing object detection. Center to our RS-Net is coarse-grained highlighting (CGH), fine-grained mining (FGM) and scale distillation (SD) modules. Through a fusion of original and complementary feature, CGH alleviates the dilemma of object semantic extraction caused by the variability of background information. Besides, FGM is designed to further enhance the feature extraction ability, which is implemented by prominent branch and ensemble branch in parallel. The former highlights the object semantics of being suppressed due to its high similarity with background. The latter mines object information as a whole to reduce the interference from noise. Finally, SD significantly improves the richness of small object information in shallow layer. Experiments conducted over DOTA and NWPU-VHR datasets confirm that the proposed RS-Net achieves competitive detection performance compared with state-of-the-art detectors.

Keywords

• Remote sensing object detection
• YOLOv4
• feature mining
• adaptive fusion