Scholarlink

<p>To address the current issues of low efficiency and unsuitability of SLM powder-spreading defect image detection algorithms for real-time monitoring deployment in industrial equipment, a lightweight SLM powder-spreading defect image segmentation algorithm called DHRA-UNet is proposed. First, an improved GhostNet (I-GN) is used as the backbone to significantly reduce the model parameters. A deep hyperparametric residual attention (DHRA) module is designed to reduce network structural information loss and enhance defect feature extraction in complex backgrounds. The skip connection structure is reconstructed, and the lightweight universal up-sampling operator Content-Aware ReAssembly of FEatures (CARAFE) and Coordinate Attention (CA) mechanism are introduced to further improve segmentation accuracy. Secondly, an SLM powder-spreading defect image acquisition system is constructed, and a mixed Illumination uniformity and Bilateral filtering (I-B) algorithm is used for image preprocessing. Finally, a weighted hybrid loss function is designed for the segmentation experiments. The results show that mIOU, Acc, and Dice metrics of our improved network are 84.72%, 96.77%, and 0.7857, respectively. The FPS is 42.2, and the model size is only 4.7% of the original model. Experiments on the VOC2007 public dataset indicate that our model has strong generalization and robustness. Overall, our improved model can be used for SLM powder-spreading defect image segmentation.</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

DHRA-UNet: a lightweight SLM powder-spreading defect image segmentation algorithm

  • 期刊论文

  • 发表日期：

    2025年8月21日

查看全文
  • ZeYuan Niu,
  • Ping Zhang,
  • Chen Zhang,
  • ZeLong Huang,
  • Xin Zhang

摘要

To address the current issues of low efficiency and unsuitability of SLM powder-spreading defect image detection algorithms for real-time monitoring deployment in industrial equipment, a lightweight SLM powder-spreading defect image segmentation algorithm called DHRA-UNet is proposed. First, an improved GhostNet (I-GN) is used as the backbone to significantly reduce the model parameters. A deep hyperparametric residual attention (DHRA) module is designed to reduce network structural information loss and enhance defect feature extraction in complex backgrounds. The skip connection structure is reconstructed, and the lightweight universal up-sampling operator Content-Aware ReAssembly of FEatures (CARAFE) and Coordinate Attention (CA) mechanism are introduced to further improve segmentation accuracy. Secondly, an SLM powder-spreading defect image acquisition system is constructed, and a mixed Illumination uniformity and Bilateral filtering (I-B) algorithm is used for image preprocessing. Finally, a weighted hybrid loss function is designed for the segmentation experiments. The results show that mIOU, Acc, and Dice metrics of our improved network are 84.72%, 96.77%, and 0.7857, respectively. The FPS is 42.2, and the model size is only 4.7% of the original model. Experiments on the VOC2007 public dataset indicate that our model has strong generalization and robustness. Overall, our improved model can be used for SLM powder-spreading defect image segmentation.