Scholarlink

<p>The identification and classification of pavement diseases are essential for road maintenance and traffic safety, yet lane markings, varying illumination and cluttered backgrounds hinder accurate diagnosis. We propose DWT-UNet-Cls, a lightweight model that embeds discrete wavelet transform (DWT) and a wavelet attention mechanism to explicitly exploit frequency information. A dual-branch wavelet-based processing strategy decomposes the input into low- and high-frequency components to capture complementary structural and textural cues, while the attention block re-weights informative frequency components through concurrent spatial and channel refinement. We evaluate the proposed method on the RDD2020 dataset and achieve a classification accuracy of 97.35% with Precision, Recall, and F1 scores outperforming classical baselines and state-of-the-art methods. We further perform ablation studies and comprehensive comparisons with classical and state-of-the-art models, and the results show that explicit wavelet-based frequency-domain modeling enables stronger robustness and more discriminative feature representations under complex pavement conditions. The source code is available at: <a href="https://github.com/L26111/DWT-UNet-Cls.">https://github.com/L26111/DWT-UNet-Cls.</a></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

Enhancing pavement disease classification through wavelet frequency features and attention mechanisms

  • 期刊论文

  • 发表日期：

    2026年5月14日

查看全文
  • Qian Liu,
  • Jinwen Wang,
  • Zheng Li,
  • Fenghua Zhu,
  • Zhaozhe Gao,
  • Benlan Shen

摘要

The identification and classification of pavement diseases are essential for road maintenance and traffic safety, yet lane markings, varying illumination and cluttered backgrounds hinder accurate diagnosis. We propose DWT-UNet-Cls, a lightweight model that embeds discrete wavelet transform (DWT) and a wavelet attention mechanism to explicitly exploit frequency information. A dual-branch wavelet-based processing strategy decomposes the input into low- and high-frequency components to capture complementary structural and textural cues, while the attention block re-weights informative frequency components through concurrent spatial and channel refinement. We evaluate the proposed method on the RDD2020 dataset and achieve a classification accuracy of 97.35% with Precision, Recall, and F1 scores outperforming classical baselines and state-of-the-art methods. We further perform ablation studies and comprehensive comparisons with classical and state-of-the-art models, and the results show that explicit wavelet-based frequency-domain modeling enables stronger robustness and more discriminative feature representations under complex pavement conditions. The source code is available at: https://github.com/L26111/DWT-UNet-Cls.