Scholarlink

<p>To address the pressing challenges of deploying deep learning models for industrial defect detection on resource-limited devices, this paper proposes a lightweight YOLOv11 model, the DSA-YOLO model, which is a lightweight model that restores accuracy through Knowledge Distillation. First, we reduce computational complexity and parameter count by replacing standard convolutions with Depthwise Separable Convolution and introducing a novel lightweight module, DSC3k2, into YOLOv11. In addition, the SimAM and the Atrous Spatial Pyramid Pooling module are employed to improve feature extraction and multi-scale detection capabilities. Finally, a two-stage cascaded Knowledge Distillation architecture based on the Knowledge Optimization model is developed to mitigate the accuracy loss resulting from lightweighting. The Knowledge Optimization model combines CNN and Transformer architectures to facilitate more effective knowledge transfer from larger models to smaller models. Experimental results on a dataset with six different defect categories show that the mAP@0.50 reaches 89.9%, demonstrating the effectiveness of the proposed method.</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

DSA-YOLO: A lightweight framework for industrial defect detection based on YOLOv11 with two-stage cascaded knowledge distillation

  • 期刊论文

  • 发表日期：

    2025年9月22日

查看全文
  • Xiaofeng Lian,
  • Xiao Sun,
  • He Liu,
  • Li Tan,
  • Jintao Li

摘要

To address the pressing challenges of deploying deep learning models for industrial defect detection on resource-limited devices, this paper proposes a lightweight YOLOv11 model, the DSA-YOLO model, which is a lightweight model that restores accuracy through Knowledge Distillation. First, we reduce computational complexity and parameter count by replacing standard convolutions with Depthwise Separable Convolution and introducing a novel lightweight module, DSC3k2, into YOLOv11. In addition, the SimAM and the Atrous Spatial Pyramid Pooling module are employed to improve feature extraction and multi-scale detection capabilities. Finally, a two-stage cascaded Knowledge Distillation architecture based on the Knowledge Optimization model is developed to mitigate the accuracy loss resulting from lightweighting. The Knowledge Optimization model combines CNN and Transformer architectures to facilitate more effective knowledge transfer from larger models to smaller models. Experimental results on a dataset with six different defect categories show that the mAP@0.50 reaches 89.9%, demonstrating the effectiveness of the proposed method.