EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

YOLOv8n-DSRS: road distress detection based on multiscale dual-path feature reorganization with selective kernel networks

Yizong Wang, Zhengrong Xiao, Xuegang An, Fei Li, Jiya Tian

PeerJ Computer Science · 03.07.2026 · 12 · e3991

DOI: 10.7717/peerj-cs.3991 ↗

Kurz gefasst

Fachartikel in der Zeitschrift PeerJ Computer Science, online erschienen am 03.07.2026. Verfasst von Yizong Wang und 4 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 33 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Natural Science Foundation of Xinjiang Uygur Autonomous Region, National College Student Innovation Training Program, Xinjiang Uygur Autonomous Region Students’ Innovation and Entrepreneurship Training Programme Projects.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Background

To address the challenges of varying target scales, complex backgrounds, and low computational efficiency in road defect detection, a road defect detection method called YOLOv8n-DSRS (You Only Look Once version 8n with DS_module, RCRep2A_FRFN, and SPPF_WD) is proposed, which is based on multi-scale dual-path feature reorganization with a selective kernel network.

Methods

First, a dynamic dual-path downsampling Dual-Scale module(DS_module) is designed, which combines spatial rearrangement strategies with standard convolutional paths in parallel to reduce spatial resolution while preserving detailed features, thereby improving detection accuracy for small-scale defects. Next, a multi-scale pooling Spatial Pyramid Pooling Fast with Weighted Downsampling (SPPF_WD) module is proposed, which combines serial max pooling with lightweight convolutions (LightConv) to enhance multi-scale feature perception capabilities. Finally, the Residual Convolutional Reparameterization with two-stage Aggregation_Feature Refinement Feed-forward Network (RCRep2A_FRFN) module was introduced, which generates redundant feature maps via Ghost Convolution (GhostConv) and combines them with the feature refinement FRFN strategy to strengthen the fusion of local and global information, thereby reducing the false negative rate.

Results

Experimental results show that the improved model YOLOv8n-DSRS achieves precision, recall, mAP (mean Average Precision) 50%, and mAP50–95% of 95.6%, 92.8%, 96.3%, and 77.3%, respectively, representing improvements of 7.4%, 12.2%, 7.8%, and 15.2% compared to the baseline model YOLOv8n. The number of parameters and Giga Floating-Point Operations Per Second (GFLOPS) are slightly lower than those of the original network model, but the overall performance is superior to other YOLOv8 series and mainstream algorithms, featuring lightweight, high accuracy, and low computational requirements. This comprehensively validates the effectiveness and superiority of this method.

Discussion

This provides strong support for applications in intelligent transportation and unmanned inspection.

Förderung

  • Natural Science Foundation of Xinjiang Uygur Autonomous Region
  • National College Student Innovation Training Program
  • Xinjiang Uygur Autonomous Region Students’ Innovation and Entrepreneurship Training Programme Projects

Zitieren

APA

Wang, Y., Xiao, Z., An, X., Li, F., & Tian, J. (2026). YOLOv8n-DSRS: road distress detection based on multiscale dual-path feature reorganization with selective kernel networks. PeerJ Computer Science, 12, e3991. https://doi.org/10.7717/peerj-cs.3991

Vancouver

Wang Y, Xiao Z, An X, Li F, Tian J. YOLOv8n-DSRS: road distress detection based on multiscale dual-path feature reorganization with selective kernel networks. PeerJ Computer Science. 2026;12:e3991. doi:10.7717/peerj-cs.3991

BibTeX
@article{Wang2026,
  author = {Wang, Yizong and Xiao, Zhengrong and An, Xuegang and Li, Fei and Tian, Jiya},
  title = {YOLOv8n-DSRS: road distress detection based on multiscale dual-path feature reorganization with selective kernel networks},
  journal = {PeerJ Computer Science},
  year = {2026},
  volume = {12},
  pages = {e3991},
  doi = {10.7717/peerj-cs.3991}
}
Kostenlos starten →