Detection of remote sensing targets with angles via modified CenterNet

•A modified CenterNet architecture guarantees the effective detection of targets with angles in remote sensing images.•A lightweight backbone possesses the capability of ensuring the high time and space efficiency.•A splicing strategy improves the detection effects for boundary targets.•Real-world experiments demonstrate the effectiveness. Remote sensing (RS) target detection aims to automatically detect and classify different targets in RS images. Due to the variability of target angles, detection of targets with angles remains a widely unsolved challenge. In this paper, we propose a modified CenterNet model to detect targets with angles in remote sensing images. The proposed network is sensitive to direction, which is beneficial to detect targets with angles precisely. In our framework, a lightweight backbone, called HRNet, is adopted instead of the traditional backbones of the conventional CenterNet, which is able to guarantee the high time and space efficiency. Moreover, a splicing strategy is introduced for optimizing the detection of boundary targets. Extensive experiments are conducted on a publicly available RS data set, and the results show that the proposed method can achieve better accuracy and higher efficiency than several state-of-the-art algorithms. [Display omitted]