Spatial hierarchy perception and hard samples metric learning for high-resolution remote sensing image object detection

Due to the different shooting angles, altitudes and scenes, remote sensing images contain many complex backgrounds and multi-scale objects. Moreover, objects in remote sensing images are much smaller relative to the backgrounds, easily occluded by buildings and trees. These cause difficult feature extraction and increase the intra-class diversity of objects, making object detection on remote sensing images more challenging. In this paper, we propose a novel remote sensing image object detection method (SHDet) based on spatial hierarchy perception component (SHPC) and hard samples metric learning (HSML). We design a SHPC to extract the feature under the different spatial hierarchies and learn the contribution weights between feature channels to enhance the feature representation. HSML is proposed to narrow the feature differences of hard samples in the same category, reducing the error detection caused by intra-class diversity. Besides, we decouple the complex background to build the pre-training datasets for pre-training the object detection model, strengthening the object feature learning. The experiments carried out on two widely used remote sensing datasets (NWPU VHR-10 and DOTA-v1.5) show that the proposed method has better detection performance compared with several state-of-the-art object detection methods.

Automated summary

A novel remote sensing image object detection method was proposed in this paper, which utilizes a spatial hierarchy perception component and hard samples metric learning to extract features under different spatial hierarchies, reduce feature differences of hard samples in the same category, and strengthen object feature learning by decoupling complex backgrounds. Experimental results demonstrate that the proposed method outperforms several state-of-the-art object detection methods in terms of detection performance.