LI-YOLOv8: Lightweight small target detection algorithm for remote sensing images that combines GSConv and PConv

Pingping Yan, Xiangming Qi and Liang Jiang

PLOS ONE, 2025, vol. 20, issue 5, 1-25

Abstract: In the domain of remote sensing image small target detection, challenges such as difficulties in extracting features of small targets, complex backgrounds that easily lead to confusion with targets, and high computational complexity with significant resource consumption are prevalent. We propose a lightweight small target detection algorithm for remote sensing images that combines GSConv and PConv, named LI-YOLOv8. Using YOLOv8n as the baseline algorithm, the activation function SiLU in the CBS at the backbone network’s SPPF is replaced with ReLU, which reduces interdependencies among parameters. Then, RFAConv is embedded after the first CBS to expand the receptive field and extract more features of small targets. An efficient Multi-Scale Attention (EMA) mechanism is embedded at the terminal of C2f within the neck network to integrate more detailed information, enhancing the focus on small targets. The head network incorporates a lightweight detection head, GP-Detect, which combines GSConv and PConv to decrease the parameter count and computational demand. Integrating Inner-IoU and Wise-IoU v3 to design the Inner-Wise IoU loss function, replacing the original CIoU loss function. This approach provides the algorithm with a gain distribution strategy, focuses on anchor boxes of ordinary quality, and strengthens generalization capability. We conducted ablation and comparative experiments on the public datasets RSOD and NWPU VHR-10. Compared to YOLOv8, our approach achieved improvements of 7.6% and 2.8% in mAP@0.5, and increases of 2.1% and 1.1% in mAP@0.5:0.95. Furthermore, Parameters and GFLOPs were reduced by 10.0% and 23.2%, respectively, indicating a significant enhancement in detection accuracy along with a substantial decrease in both parameters and computational costs. Generalization experiments were conducted on the TinyPerson, LEVIR-ship, brain-tumor, and smoke_fire_1 datasets. The mAP@0.5 metric improved by 2.6%, 5.3%, 2.6%, and 2.3%, respectively, demonstrating the algorithm’s robust performance.

Date: 2025
References: View complete reference list from CitEc
Citations:

Downloads: (external link)
https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0321026 (text/html)
https://journals.plos.org/plosone/article/file?id= ... 21026&type=printable (application/pdf)

Related works:
This item may be available elsewhere in EconPapers: Search for items with the same title.

Export reference: BibTeX RIS (EndNote, ProCite, RefMan) HTML/Text

Persistent link: https://EconPapers.repec.org/RePEc:plo:pone00:0321026

DOI: 10.1371/journal.pone.0321026

Access Statistics for this article

More articles in PLOS ONE from Public Library of Science
Bibliographic data for series maintained by plosone ().