Scholarlink

<p>This paper presents the development of a deep learning model specifically designed for error detection in the assembly of electronic components on a printed circuit board. The proposed Convolutional Deep Learning Model (CDLM) consists of three stages: segmentation, feature extraction and classification. The segmentation is based on the model for detecting objects in images: Mask r-cnn. The feature extraction stage utilizes Convolutional Neural Networks and follows an approach inspired by human vision—detecting first coarse structures, then medium details, and finally fine details. The classification stage is performed using Fully Connected (FC) neural networks. For training and evaluation, a proprietary dataset was created, consisting of 2376 images representing five types of defects on the boards, plus the corrected one; all boards type are equally represented. Finally, the model was implemented on a computer that processes digital images of the boards to provide real-time inspection. The proposal is able to determine whether or not a board has one of the five proposed defects. The results show a zero error and 100% accuracy in classifying the boards.</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

Deep learning model for automated visual inspection of electronic boards

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月8日

查看全文
  • José Antonio Lara-Chávez,
  • Carlos Avilés-Cruz,
  • Miguel Magos-Rivera

摘要

This paper presents the development of a deep learning model specifically designed for error detection in the assembly of electronic components on a printed circuit board. The proposed Convolutional Deep Learning Model (CDLM) consists of three stages: segmentation, feature extraction and classification. The segmentation is based on the model for detecting objects in images: Mask r-cnn. The feature extraction stage utilizes Convolutional Neural Networks and follows an approach inspired by human vision—detecting first coarse structures, then medium details, and finally fine details. The classification stage is performed using Fully Connected (FC) neural networks. For training and evaluation, a proprietary dataset was created, consisting of 2376 images representing five types of defects on the boards, plus the corrected one; all boards type are equally represented. Finally, the model was implemented on a computer that processes digital images of the boards to provide real-time inspection. The proposal is able to determine whether or not a board has one of the five proposed defects. The results show a zero error and 100% accuracy in classifying the boards.