Scholarlink

<p>Bridge defect detection is a critical process for ensuring structural safety, and with the continuous development of computer vision techniques, its applications in bridge defect recognition have become increasingly widespread. However, bridge inspection scenarios are generally characterized by limited computational resources, small defect targets, and high inter-class similarity, which pose significant challenges to both lightweight model deployment and recognition accuracy. To address these issues, this study proposes a two-stage object recognition framework enhanced with pixel-level segmentation, specifically tailored for resource-constrained mobile platforms such as unmanned aerial vehicles (UAVs). In the first stage, a lightweight YOLOv11n model is constructed based on the MobileNetV4 architecture, where a Universal Inverted Bottleneck replaces conventional convolutional modules. This design achieves 53% model compression&#xa0;and significantly improves inference efficiency while maintaining high detection accuracy (94.9% mAP50 in single-class tasks). In the second stage, a Vision Transformer&#xa0;serves as a fine-grained classifier to resolve low-confidence detections and disambiguate highly similar defect categories, boosting classification robustness. Furthermore, to enable precise geometric quantification of defects—such as crack length and spalling area—an&#xa0;enhanced U-Net segmentation head&#xa0;is integrated downstream of the detector. This module incorporates residual connections, SiLU activation, and attention-guided skip connections, achieving a Dice coefficient of 0.854 and mAP of 81.22% on ROI-cropped images. Additionally, background images are incorporated during training to strengthen the model’s anti-interference capability across all stages. Experimental results demonstrate that, compared with the original YOLOv11n baseline, the proposed framework improves fine-grained defect classification accuracy by 4.7% and increases the overall F1-score by approximately 10%, while simultaneously enabling accurate pixel-wise defect delineation. This study provides a practical and efficient solution for&#xa0;joint detection, classification, and segmentation&#xa0;in real-world bridge defect inspection under stringent computational constraints.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

UIB-Enhanced YOLOv11 Coupled with Vision Transformer: A Two-Stage Architecture for Efficient Bridge Defect Recognition

  • 期刊论文

  • 发表日期：

    2026年7月18日

查看全文
  • Yucheng Lu,
  • Ge Zhang,
  • Shuai Teng,
  • Gongfa Chen,
  • Xiaomei Yang

摘要

Bridge defect detection is a critical process for ensuring structural safety, and with the continuous development of computer vision techniques, its applications in bridge defect recognition have become increasingly widespread. However, bridge inspection scenarios are generally characterized by limited computational resources, small defect targets, and high inter-class similarity, which pose significant challenges to both lightweight model deployment and recognition accuracy. To address these issues, this study proposes a two-stage object recognition framework enhanced with pixel-level segmentation, specifically tailored for resource-constrained mobile platforms such as unmanned aerial vehicles (UAVs). In the first stage, a lightweight YOLOv11n model is constructed based on the MobileNetV4 architecture, where a Universal Inverted Bottleneck replaces conventional convolutional modules. This design achieves 53% model compression and significantly improves inference efficiency while maintaining high detection accuracy (94.9% mAP50 in single-class tasks). In the second stage, a Vision Transformer serves as a fine-grained classifier to resolve low-confidence detections and disambiguate highly similar defect categories, boosting classification robustness. Furthermore, to enable precise geometric quantification of defects—such as crack length and spalling area—an enhanced U-Net segmentation head is integrated downstream of the detector. This module incorporates residual connections, SiLU activation, and attention-guided skip connections, achieving a Dice coefficient of 0.854 and mAP of 81.22% on ROI-cropped images. Additionally, background images are incorporated during training to strengthen the model’s anti-interference capability across all stages. Experimental results demonstrate that, compared with the original YOLOv11n baseline, the proposed framework improves fine-grained defect classification accuracy by 4.7% and increases the overall F1-score by approximately 10%, while simultaneously enabling accurate pixel-wise defect delineation. This study provides a practical and efficient solution for joint detection, classification, and segmentation in real-world bridge defect inspection under stringent computational constraints.