Scholarlink

Abstract <p>This study analyzes binary image classification techniques adapted for detecting weld defects under conditions of limited data availability. A thorough analysis of methodologies including augmentation [4], hard mining [6], transfer learning [3], and training with a small number of training examples has been conducted, examining their effectiveness in scenarios with limited training data. The results of this rigorous analysis served as the basis for the development of a model training method called task switching. This innovative approach has demonstrated significant improvements in object classification accuracy, particularly when working with small amounts of data. These improvements are due to the unique characteristics of the receptive fields of the selected models as a result of the transition to detection. The task switching methodology demonstrates the ability to exploit these characteristics, resulting in a significant increase in the accuracy of defect identification in binary image classification. The fusion of task switching and advanced detection mechanisms creates a paradigm shift, significantly increasing the accuracy of weld defect detection, despite the limited data availability.</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

Welding Defect Detection Using Small Data

  • 期刊论文

  • 发表日期：

    2025年12月1日

查看全文
  • Anton Evgenievich Misnik,
  • Eduard Aleksandrovich Kapelko

摘要

Abstract

This study analyzes binary image classification techniques adapted for detecting weld defects under conditions of limited data availability. A thorough analysis of methodologies including augmentation [4], hard mining [6], transfer learning [3], and training with a small number of training examples has been conducted, examining their effectiveness in scenarios with limited training data. The results of this rigorous analysis served as the basis for the development of a model training method called task switching. This innovative approach has demonstrated significant improvements in object classification accuracy, particularly when working with small amounts of data. These improvements are due to the unique characteristics of the receptive fields of the selected models as a result of the transition to detection. The task switching methodology demonstrates the ability to exploit these characteristics, resulting in a significant increase in the accuracy of defect identification in binary image classification. The fusion of task switching and advanced detection mechanisms creates a paradigm shift, significantly increasing the accuracy of weld defect detection, despite the limited data availability.