Scholarlink

Abstract <p>Although deep neural networks have made significant advancements in machine vision, detecting small defects on thin films using deep neural networks remains a challenging task. The shape and size of the target defect are often uncertain, while environmental conditions such as illumination and camera vibration can affect image quality. Additionally, the defect detection model needs to process images at high speed, further adding to the complexity of the task. To address these challenges, this paper proposes a novel approach by utilizing the “one-stage” object detection method as the backbone structure, while also enhancing global representation and multilevel feature fusion for improved defect detection. Evaluation results demonstrate that the proposed network structure not only performs well on our datasets but also outperforms other state-of-the-art object detection methods when tested on public datasets.</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

Accurate Thin Film Defect Detection Using Transformer and Guided Feature Fusion

  • 期刊论文

  • 发表日期：

    2025年6月25日

查看全文
  • Tao Yang,
  • Qinghuan Liu,
  • Huiyan Wang,
  • Shengjie Li

摘要

Abstract

Although deep neural networks have made significant advancements in machine vision, detecting small defects on thin films using deep neural networks remains a challenging task. The shape and size of the target defect are often uncertain, while environmental conditions such as illumination and camera vibration can affect image quality. Additionally, the defect detection model needs to process images at high speed, further adding to the complexity of the task. To address these challenges, this paper proposes a novel approach by utilizing the “one-stage” object detection method as the backbone structure, while also enhancing global representation and multilevel feature fusion for improved defect detection. Evaluation results demonstrate that the proposed network structure not only performs well on our datasets but also outperforms other state-of-the-art object detection methods when tested on public datasets.