Key points are not available for this paper at this time.
Abstract Due to the complex backgrounds of remote sens2 ing images and the small size of aircraft targets, the results of 3 commonly used detection algorithms in small aircraft target 4 detection are not satisfactory. Moreover, the prevalent deep 5 learning algorithms are generally too cumbersome to adapt 6 to the resource-constrained satellite platforms. To improve 7 detection accuracy while maintaining model simplicity for 8 satellite on-orbit small aircraft detection, we develop an im9 proved algorithm based on YOLOv5, called LEN-YOLO. 10 Firstly, we adopt the EIoU Loss for target localization, en11 abling the network to effectively focus on small aircraft tar12 gets. Second, a Lite backbone is designed by discarding high 13 semantic information, using low semantic feature maps to 14 detect small targets. Finally, we propose a BSG-FPN struc15 ture to fuse feature maps of different scales to increase de16 tailed information. Experimental results on RSOD and DIOR 17 datasets demonstrate compared to the baseline YOLOv5, 18 LEN-YOLO achieves an increase of 5.1% and 4.2% in APs 19 respectively. Notably, parameters are reduced by 78.3% and 20 floating-point operations by 33.2%.
Wu et al. (Thu,) studied this question.