Scholarlink

<p>Image defect detection of power transmission and transformation equipment is a key technology in the operation and maintenance of power systems, which is of vital importance for ensuring the safe and stable operation of the power grid and enhancing its self-healing capacity. However, this task has long been confronted with the core challenge of a limited number of defect samples for specific categories. Traditional deep learning models trained on large-scale data find it difficult to achieve accurate detection under small-sample conditions. To address this issue, this study proposes a lightweight segmentation network based on the Mamba architecture, Residual Mamba (ResMamba). This network adopts a six-level U-shaped codec structure and innovatively integrates the Mamba module into the visual state space (VSS) as the codec link. Its core ResVSS module significantly reduces the number of parameters by removing a redundant linear layer (accounting for 38% of the original VSS module parameters) in the internal shortcut connection of the original VSS module, and introduces deep convolutional blocks and learnable scale parameters to dynamically scale residual connections. It enhances feature representation ability while reducing model complexity. In addition, the skip connection part introduces a multi-level and multi-scale information fusion mechanism, effectively integrating cross-scale features by generating spatial and channel attention maps, thereby enhancing the model’s performance in multi-specification defect detection. Experimental results on public datasets (including specialized small-sample settings) show that ResMamba achieves superior segmentation accuracy while maintaining a low parameter count, effectively balancing computational efficiency and detection performance. It outperforms both general segmentation models and domain-specific models for power equipment defect detection, providing a reliable new solution for small-sample defect detection of power equipment and enhancing the self-healing ability of power grids.</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

Study on mmage defect recognition and classification of power transmission equipment based on lightweight model residual Mamba

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月9日

查看全文
  • Cui Jindong,
  • Song Weijie,
  • Guan Shan,
  • Sun Liang,
  • Tian Hongliang,
  • Wang Fuwang

摘要

Image defect detection of power transmission and transformation equipment is a key technology in the operation and maintenance of power systems, which is of vital importance for ensuring the safe and stable operation of the power grid and enhancing its self-healing capacity. However, this task has long been confronted with the core challenge of a limited number of defect samples for specific categories. Traditional deep learning models trained on large-scale data find it difficult to achieve accurate detection under small-sample conditions. To address this issue, this study proposes a lightweight segmentation network based on the Mamba architecture, Residual Mamba (ResMamba). This network adopts a six-level U-shaped codec structure and innovatively integrates the Mamba module into the visual state space (VSS) as the codec link. Its core ResVSS module significantly reduces the number of parameters by removing a redundant linear layer (accounting for 38% of the original VSS module parameters) in the internal shortcut connection of the original VSS module, and introduces deep convolutional blocks and learnable scale parameters to dynamically scale residual connections. It enhances feature representation ability while reducing model complexity. In addition, the skip connection part introduces a multi-level and multi-scale information fusion mechanism, effectively integrating cross-scale features by generating spatial and channel attention maps, thereby enhancing the model’s performance in multi-specification defect detection. Experimental results on public datasets (including specialized small-sample settings) show that ResMamba achieves superior segmentation accuracy while maintaining a low parameter count, effectively balancing computational efficiency and detection performance. It outperforms both general segmentation models and domain-specific models for power equipment defect detection, providing a reliable new solution for small-sample defect detection of power equipment and enhancing the self-healing ability of power grids.