Scholarlink

<p>Weld defect detection plays an important role in ensuring the quality and safety of welded structures and the reliability of products.&#xa0;Current deep learning-based models for weld defect detection have significantly enhanced the accuracy of identifying such defects. Nevertheless, these models tend to be overly large and possess limited feature extraction capabilities, resulting in inefficient detection. This paper proposed an accurate and lightweight YOLOv8-DPE algorithm based on the You Only Look Once (YOLO) framework. The proposes algorithm introduces a novel and efficient multi-scale feature extraction Dilation-Wise Residual (DWR) module in the backbone network of feature extraction, which makes efficient use of depth-extended convolution and reduces the difficulty of extracting image feature information. The application of the Poly&#xa0;Kernel Inception (PKI) block with a contextual anchor attention mechanism in the C2f module can better extract and fuse global features. Furthermore, introduced Partial Convolution (PConv) in the Detection Head to make the model optimized to be lightweight. Finally, comparative experiments and ablation experiments on the self-constructed weld defect dataset and NEU-DET dataset verified the effectiveness and superiority of the proposed algorithm. Experimental results indicate that compared to the YOLOv8n model, YOLOv8-DPE achieves 99.0% and 81.6% accuracy in mAP50 and mAP50:95, respectively, marking improvements of 1.5% and 4.6% over the original algorithm, while the FLOPs, number of parameters and model size were reduced by 38.27%, 29.55%, and 27.0%, respectively. Compared to other popular object detection algorithms, the proposed improved algorithm performs better overall.</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

YOLOv8-DPE: a novel lightweight object detection network for weld defect detection

  • 期刊论文

  • 发表日期：

    2025年8月29日

查看全文
  • Lihua Yang,
  • Yi Gu,
  • Hao Feng

摘要

Weld defect detection plays an important role in ensuring the quality and safety of welded structures and the reliability of products. Current deep learning-based models for weld defect detection have significantly enhanced the accuracy of identifying such defects. Nevertheless, these models tend to be overly large and possess limited feature extraction capabilities, resulting in inefficient detection. This paper proposed an accurate and lightweight YOLOv8-DPE algorithm based on the You Only Look Once (YOLO) framework. The proposes algorithm introduces a novel and efficient multi-scale feature extraction Dilation-Wise Residual (DWR) module in the backbone network of feature extraction, which makes efficient use of depth-extended convolution and reduces the difficulty of extracting image feature information. The application of the Poly Kernel Inception (PKI) block with a contextual anchor attention mechanism in the C2f module can better extract and fuse global features. Furthermore, introduced Partial Convolution (PConv) in the Detection Head to make the model optimized to be lightweight. Finally, comparative experiments and ablation experiments on the self-constructed weld defect dataset and NEU-DET dataset verified the effectiveness and superiority of the proposed algorithm. Experimental results indicate that compared to the YOLOv8n model, YOLOv8-DPE achieves 99.0% and 81.6% accuracy in mAP50 and mAP50:95, respectively, marking improvements of 1.5% and 4.6% over the original algorithm, while the FLOPs, number of parameters and model size were reduced by 38.27%, 29.55%, and 27.0%, respectively. Compared to other popular object detection algorithms, the proposed improved algorithm performs better overall.