Scholarlink

<p>To address the issue of poor accuracy in cold-rolled sheet defect detection, a deep learning model enhanced with omni-dimensional dynamic and global attention mechanisms is proposed. First, omni-dimensional dynamic and global attention mechanisms are integrated into the backbone network. The omni-dimensional dynamic mechanism employs a multi-dimensional attention strategy through parallelism, learning complementary attention along four dimensions: spatial size, input channels, output channels, and the number of convolution kernels. This effectively enhances the network’s feature extraction capability and improves its understanding and modeling of input data. The global attention mechanism captures all significant features across the channel and spatial dimensions, considering the interactions among these dimensions, thereby reducing information loss and enhancing global dimension interaction features. This leads to more comprehensive defect feature extraction, improving detection accuracy and stability. Second, we collected images of ten common types of cold-rolled sheet defects from a specific cold-rolling mill and expanded the dataset using image augmentation methods. Ablation experiments were conducted under the established cold-rolled defect dataset to compare our model with three mainstream attention mechanisms. Additionally, comparative experiments with classic two-stage models, one-stage models, and the latest state-of-the-art models were performed to validate the effectiveness of the proposed model. Attention heatmap methods were used to visualize the feature focus areas, enhancing the model’s interpretability. Finally, the proposed model was validated and tested on the ten types of cold-rolled sheet defects. Results show that our model achieved a mean average precision of 97.53&#xa0;pct, outperforming the comparison models. The average inference time per image was only 39.5&#xa0;ms, and the model complexity was relatively low, meeting the real-time and applicability requirements of actual production. Moreover, the proposed model significantly outperformed the three comparison attention mechanisms in multi-category feature attention, significantly improving defect feature extraction and detection accuracy, thereby providing new technical support for intelligent defect detection in cold-rolled sheets.</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

Deep Learning Model for Cold-Rolled Plate Defect Detection Based on Omni-dimensional Dynamic Convolution and Global Attention Mechanism Enhancement

  • 期刊论文

  • 发表日期：

    2025年6月4日

查看全文
  • Wenguang Xu,
  • Jianhua Liu,
  • Yang He,
  • Yuanmei Yang,
  • Xin Xie,
  • Xiaodong Yang

摘要

To address the issue of poor accuracy in cold-rolled sheet defect detection, a deep learning model enhanced with omni-dimensional dynamic and global attention mechanisms is proposed. First, omni-dimensional dynamic and global attention mechanisms are integrated into the backbone network. The omni-dimensional dynamic mechanism employs a multi-dimensional attention strategy through parallelism, learning complementary attention along four dimensions: spatial size, input channels, output channels, and the number of convolution kernels. This effectively enhances the network’s feature extraction capability and improves its understanding and modeling of input data. The global attention mechanism captures all significant features across the channel and spatial dimensions, considering the interactions among these dimensions, thereby reducing information loss and enhancing global dimension interaction features. This leads to more comprehensive defect feature extraction, improving detection accuracy and stability. Second, we collected images of ten common types of cold-rolled sheet defects from a specific cold-rolling mill and expanded the dataset using image augmentation methods. Ablation experiments were conducted under the established cold-rolled defect dataset to compare our model with three mainstream attention mechanisms. Additionally, comparative experiments with classic two-stage models, one-stage models, and the latest state-of-the-art models were performed to validate the effectiveness of the proposed model. Attention heatmap methods were used to visualize the feature focus areas, enhancing the model’s interpretability. Finally, the proposed model was validated and tested on the ten types of cold-rolled sheet defects. Results show that our model achieved a mean average precision of 97.53 pct, outperforming the comparison models. The average inference time per image was only 39.5 ms, and the model complexity was relatively low, meeting the real-time and applicability requirements of actual production. Moreover, the proposed model significantly outperformed the three comparison attention mechanisms in multi-category feature attention, significantly improving defect feature extraction and detection accuracy, thereby providing new technical support for intelligent defect detection in cold-rolled sheets.