Scholarlink

<p>Object detection is a crucial task in computer vision, aiming to identify and locate specific objects in images or videos. Despite significant advancements, existing object detection models struggle with defect detection accuracy, especially for the detection of small defects and complex backgrounds. We propose GFD-Net, a real-time defect detection algorithm that leverages Global Convolutional Additive Self-attention (GCAS) and a Fully-aggregated Frequency-aware Feature Pyramid Network (FF-FPN). GCAS enhances spatial feature learning by integrating local and global contextual information, while FF-FPN ensures efficient multi-scale feature fusion and information transmission. Furthermore, we designed a Dual Detection Heads with Shared Convolutions (DHSC) structure to optimize classification and localization tasks. Experiments on a self-built Micro-Electrical Connector Defect (MECD) dataset, GFD-Net achieves a precision of 97% with only 2MB parameters, showcasing its efficiency and accuracy for real-time object detection. On the NEU-DET dataset demonstrate that GFD-Net achieves a mean Average Precision (mAP) of 77.6%, and the number of parameters is reduced by approximately 28%. The code can be found at: <a href="https://github.com/ZJWstar/GFD-Net.git">https://github.com/ZJWstar/GFD-Net.git</a></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

GFD-Net: Micro-electrical connector defect localization network with global additive attention and frequency-aware feature fusion

  • 期刊论文

  • 发表日期：

    2025年7月22日

查看全文
  • Hongqi Wang,
  • Jiawen Zhao,
  • Qunpo Liu,
  • Naohiko Hanajima,
  • Xuhui Bu

摘要

Object detection is a crucial task in computer vision, aiming to identify and locate specific objects in images or videos. Despite significant advancements, existing object detection models struggle with defect detection accuracy, especially for the detection of small defects and complex backgrounds. We propose GFD-Net, a real-time defect detection algorithm that leverages Global Convolutional Additive Self-attention (GCAS) and a Fully-aggregated Frequency-aware Feature Pyramid Network (FF-FPN). GCAS enhances spatial feature learning by integrating local and global contextual information, while FF-FPN ensures efficient multi-scale feature fusion and information transmission. Furthermore, we designed a Dual Detection Heads with Shared Convolutions (DHSC) structure to optimize classification and localization tasks. Experiments on a self-built Micro-Electrical Connector Defect (MECD) dataset, GFD-Net achieves a precision of 97% with only 2MB parameters, showcasing its efficiency and accuracy for real-time object detection. On the NEU-DET dataset demonstrate that GFD-Net achieves a mean Average Precision (mAP) of 77.6%, and the number of parameters is reduced by approximately 28%. The code can be found at: https://github.com/ZJWstar/GFD-Net.git