LEN-YOLO: a lightweight remote sensing small aircraft object detection model for satellite on-orbit detection

The performance of conventional detection algorithms in small aircraft target detection is often unsatisfactory due to the intricate backgrounds of remote sensing images and the diminutive size of aircraft targets. Furthermore, prevalent deep learning algorithms typically prove overly complex for integration into resource-constrained satellite platforms. In response to these challenges, an enhanced algorithm named LEN-YOLO (Lite backbone - Enhanced Neck - YOLO) has been devised to enhance detection accuracy while preserving model simplicity for the detection of small aircraft in satellite on-orbit scenarios. First, the EIoU Loss is adopted for target localization, enabling the network to effectively focus on small aircraft targets. Second, a Lite backbone is designed by discarding high semantic information, using low-semantic feature maps to detect small targets. Finally, a Bidirectional Weighted FPN based on SimAM and GSConv (BSG-FPN) is proposed to fuse feature maps of different scales to increase detailed information. Experimental results on RSOD and DIOR datasets demonstrate compared to the baseline YOLOv5, LEN-YOLO achieves an increase of 5.1% and 4.2% in APs\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_s$$\end{document} respectively. Notably, parameters are reduced by 78.3% and floating-point operations by 33.2%.