Scholarlink

<p>With the increasing demand for road maintenance, traditional manual classification and conventional image processing methods are limited by low efficiency and poor robustness in complex environments. Although Convolutional Neural Networks (CNNs) have improved performance, they often struggle to capture long-range dependencies and remain sensitive to environmental interference. To address these challenges, this study presents a task-driven improved Swin-T framework for road damage classification. The framework integrates three key components: a Locality-Preserving and Cross-scale Enhancement (LPCE) mechanism to enhance local spatial continuity, a Feature Alignment Module (FAM) to improve spatial–semantic consistency with low computational overhead, and a Window Size Modification (WSM) strategy to better handle high-aspect-ratio defects. Experimental results on the RDD2022 (China_Motorbike) dataset demonstrate that the proposed framework achieves competitive performance, reaching 97.0% accuracy with only minor increases in parameters (2.5%) and FLOPs (2.7%). In addition, the model maintains real-time inference speed and exhibits stable performance under controlled low-light and noise conditions, indicating improved robustness for road classification tasks.</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

Road defect classification based on improved Swin-T

  • 期刊论文

  • 发表日期：

    2026年6月29日

查看全文
  • Yachao Si,
  • Yi Zhang,
  • Mingzhan Zhao,
  • Lijuan Li

摘要

With the increasing demand for road maintenance, traditional manual classification and conventional image processing methods are limited by low efficiency and poor robustness in complex environments. Although Convolutional Neural Networks (CNNs) have improved performance, they often struggle to capture long-range dependencies and remain sensitive to environmental interference. To address these challenges, this study presents a task-driven improved Swin-T framework for road damage classification. The framework integrates three key components: a Locality-Preserving and Cross-scale Enhancement (LPCE) mechanism to enhance local spatial continuity, a Feature Alignment Module (FAM) to improve spatial–semantic consistency with low computational overhead, and a Window Size Modification (WSM) strategy to better handle high-aspect-ratio defects. Experimental results on the RDD2022 (China_Motorbike) dataset demonstrate that the proposed framework achieves competitive performance, reaching 97.0% accuracy with only minor increases in parameters (2.5%) and FLOPs (2.7%). In addition, the model maintains real-time inference speed and exhibits stable performance under controlled low-light and noise conditions, indicating improved robustness for road classification tasks.