A context-scale-aware detector and a new benchmark for remote sensing small weak object detection in unmanned aerial vehicle images

• Published in: International journal of applied earth observation and geoinformation 2022-08, Vol.112, p.102966, Article 102966
• Main Authors: Han, Wei, Li, Jun, Wang, Sheng, Wang, Yi, Yan, Jining, Fan, Runyu, Zhang, Xiaohan, Wang, Lizhe
• Format: Article
• Language: English
• Subjects: High-resolution remote sensing; Object detection; Remote sensing; Unmanned aerial vehicle
• ISSN: 1569-8432; 1872-826X
• DOI: 10.1016/j.jag.2022.102966

Due to the fast development of imaging sensors used in remote sensing, high-resolution images can increasingly exhibit fine-grained information on the Earth’s surface, which makes detecting real-world small-scale, weak-feature-response geospatial targets possible. Detecting these small weak targets is of great significance in some applications. Although remarkable efforts have been made to develop small weak object detection, existing works mainly focus on processing satellite images. Unmanned aerial vehicle (UAV) remote sensing images have very high resolutions and short revisit times, thus making them a novel kind of data for Earth observation, and are more rarely studied. Meanwhile, due to the characteristics of UAV images, some problems, such as high intra-class variations, complex contexts, and noise, are pronounced. To promote the remote sensing detection of small weak objects in UAV images, this paper proposes a 10-category UAV object detection dataset, namely UAVOD-10, 11The dataset is available at https://github.com/weihancug/10-category-UAV-small-weak-object-detection-dataset-UAVOD10. and a novel context-scale-aware detector, namely, CSADet. The objects in UAVOD-10 are sufficiently varied and affected by imaging conditions and their contexts. Some of them, such as cable towers, wells, and cultivation mesh cages, have small scales and weak feature responses, making them difficult to recognize. To locate these kinds of objects efficiently, CSADet first utilizes a context-aware module that can jointly explore valuable local and global contexts and then applies a multi-scale feature refinement module to extract and share all the levels of the features. Extensive experiments on the proposed UAVOD-10 dataset demonstrate its remarkable detection performance. •A real-world UAV object detection dataset with sufficient variations is presented.•A CSADet is developed to learn the objects’ context and discriminative features.•The superiority of the CSADet in small, weak object detection is proven.