Scholarlink

<p>To address the technical challenges such as incomplete laser stripe segmentation and low detection accuracy caused by arc light interference in complex welding environments, a robust laser stripe extraction method based on a REA-UNet algorithmn is proposed. Building upon the U-Net architecture, the proposed framework integrates residual modules into the encoder to improve gradient propagation and feature preservation. A multi-branch downsampling structure is designed for fusing multi-scale edge features to help reduce information loss. At the encoder’s end, an Improved Atrous Spatial Pyramid Pooling (IASPP) module is incorporated to expand the receptive field and aimed at enhancing multi-scale feature extraction. Furthermore, an Efficient Channel Attention (ECA) mechanism is integrated during the decoder stage to emphasize critical feature channels, while depthwise separable convolutions are implemented to improve computational efficiency. A hybrid loss function is also adopted to address class imbalance commonly observed in laser stripe regions. Experimental results demonstrate that the proposed REA-UNet model achieves 94.48% in Mean Pixel Accuracy (<i>MPA</i>), 87.46% in Mean Intersection over Union (<i>MIoU</i>), and 85.96% in <i>Dice</i> coefficient, outperform other comparison algorithms on the tested dataset. While maintaining high precision, the model’s average inference time per image is controlled at 25.21 ms, suggesting potential for real-time application in welding tasks. These results suggest that the method provides improved robustness and resistance to arc-related interference in weld seam laser stripe segmentation, indicating its potential applicability in automated arc welding systems.</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

Laser stripe extraction for weld seams based on a REA-UNet algorithm

  • 期刊论文

  • 发表日期：

    2025年10月28日

查看全文
  • Chang Jiang,
  • Haonan Wu,
  • Jiatong Zhu,
  • Huiming Wu,
  • Dongtao Xu,
  • Zhendong Song

摘要

To address the technical challenges such as incomplete laser stripe segmentation and low detection accuracy caused by arc light interference in complex welding environments, a robust laser stripe extraction method based on a REA-UNet algorithmn is proposed. Building upon the U-Net architecture, the proposed framework integrates residual modules into the encoder to improve gradient propagation and feature preservation. A multi-branch downsampling structure is designed for fusing multi-scale edge features to help reduce information loss. At the encoder’s end, an Improved Atrous Spatial Pyramid Pooling (IASPP) module is incorporated to expand the receptive field and aimed at enhancing multi-scale feature extraction. Furthermore, an Efficient Channel Attention (ECA) mechanism is integrated during the decoder stage to emphasize critical feature channels, while depthwise separable convolutions are implemented to improve computational efficiency. A hybrid loss function is also adopted to address class imbalance commonly observed in laser stripe regions. Experimental results demonstrate that the proposed REA-UNet model achieves 94.48% in Mean Pixel Accuracy (MPA), 87.46% in Mean Intersection over Union (MIoU), and 85.96% in Dice coefficient, outperform other comparison algorithms on the tested dataset. While maintaining high precision, the model’s average inference time per image is controlled at 25.21 ms, suggesting potential for real-time application in welding tasks. These results suggest that the method provides improved robustness and resistance to arc-related interference in weld seam laser stripe segmentation, indicating its potential applicability in automated arc welding systems.