Scholarlink

<p>Deep learning-based image segmentation techniques have emerged as a leading approach for detecting industrial defects. However, training deep learning models necessitates a substantial number of defect samples, and the scarcity of negative industrial defect samples complicates their implementation. Additionally, in real industrial environments, defects are often coupled in complex background patterns, making them difficult to identify. To tackle these issues, we propose a Gaussian Feature Augmented Few-Shot Model (GF-Few) for real-time industrial defect detection tasks. By leveraging the generalization ability of the few-shot network structure for images with a limited number of support sets, we introduce an Adaptive Gaussian Membership Feature Module (GM-FE) and a specialized Gaussian Feature Loss Function (GFloss) to enhance the model’s feature extraction process. This combination uses a Gaussian function to adjust the distribution range of the model in the feature space, thereby improving the feature representation of the image. Furthermore, we design a Defect Prototype Alignment Module (DPAM) to compare the differences in defect features between the support set and the query set, which facilitates defect detection by employing a simpler network architecture. This method fully utilizes prior knowledge from a limited number of support sets, enabling more efficient real-time industrial defect detection. We selected three public industrial datasets and collected a Printing-Packaging-Box dataset to evaluate the model. The experimental results demonstrate that the method strikes a balance between detection accuracy and model size, achieving a detection accuracy of 92.21%, with a parameter count of 9.37&#xa0;M and a detection speed of 183 frames per second, effectively meeting the requirements for industrial environments.</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

GF-few: a real-time lightweight Gaussian feature enhanced few-shot network for industrial defect detection

  • 期刊论文

  • 发表日期：

    2025年8月29日

查看全文
  • Qiurui Ma,
  • Erhu Zhang,
  • Yajun Chen,
  • Jinghong Duan

摘要

Deep learning-based image segmentation techniques have emerged as a leading approach for detecting industrial defects. However, training deep learning models necessitates a substantial number of defect samples, and the scarcity of negative industrial defect samples complicates their implementation. Additionally, in real industrial environments, defects are often coupled in complex background patterns, making them difficult to identify. To tackle these issues, we propose a Gaussian Feature Augmented Few-Shot Model (GF-Few) for real-time industrial defect detection tasks. By leveraging the generalization ability of the few-shot network structure for images with a limited number of support sets, we introduce an Adaptive Gaussian Membership Feature Module (GM-FE) and a specialized Gaussian Feature Loss Function (GFloss) to enhance the model’s feature extraction process. This combination uses a Gaussian function to adjust the distribution range of the model in the feature space, thereby improving the feature representation of the image. Furthermore, we design a Defect Prototype Alignment Module (DPAM) to compare the differences in defect features between the support set and the query set, which facilitates defect detection by employing a simpler network architecture. This method fully utilizes prior knowledge from a limited number of support sets, enabling more efficient real-time industrial defect detection. We selected three public industrial datasets and collected a Printing-Packaging-Box dataset to evaluate the model. The experimental results demonstrate that the method strikes a balance between detection accuracy and model size, achieving a detection accuracy of 92.21%, with a parameter count of 9.37 M and a detection speed of 183 frames per second, effectively meeting the requirements for industrial environments.