Scholarlink

<p>The rich and complex texture information of industrial products, and the fact that the number of normal products often far exceeds the number of defective products in industrial scenarios, poses a great challenge to product quality inspection. In order to solve this challenge, a weakly supervised defect detection model based on nested U-Net was proposed: the nested U-Net was used as the main body of the model, and an attention module was introduced, which could obtain the relationship between local features and between feature channels. Furthermore, A weakly supervised training strategy was employed: the defect mask from the Berlin noise, external data and normal samples are used to synthesize the defects, and a few real defect samples are randomly inserted into the synthetic defect samples to train the detection model. Experimental validation was carried out on the public datasets MVTec AD, DAGM, MT and custom CT (computed tomography) composite material dataset, and the evaluation indicators included image-level AUC (area under the receiver’s operating characteristic curve), pixel-level AUC and AP (average accuracy). Experimental results show that the proposed method achieves excellent performance of 99.9%/98.7%/84.1%, 99.1%/95.3%/76.1%, 100%/98.1%/86.7% and 73.6%/69.1%/36.0% on three types of metrics on four datasets, respectively, which is better than the current advanced model.</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

Weakly Supervised Complex Texture Defect Detection Based on Nested U-Net Architecture

  • 期刊论文

  • 发表日期：

    2025年2月9日

查看全文
  • Kuidong Huang,
  • Yang Zeng,
  • Julong Zhao,
  • Shijie Chai,
  • Fuqiang Yang

摘要

The rich and complex texture information of industrial products, and the fact that the number of normal products often far exceeds the number of defective products in industrial scenarios, poses a great challenge to product quality inspection. In order to solve this challenge, a weakly supervised defect detection model based on nested U-Net was proposed: the nested U-Net was used as the main body of the model, and an attention module was introduced, which could obtain the relationship between local features and between feature channels. Furthermore, A weakly supervised training strategy was employed: the defect mask from the Berlin noise, external data and normal samples are used to synthesize the defects, and a few real defect samples are randomly inserted into the synthetic defect samples to train the detection model. Experimental validation was carried out on the public datasets MVTec AD, DAGM, MT and custom CT (computed tomography) composite material dataset, and the evaluation indicators included image-level AUC (area under the receiver’s operating characteristic curve), pixel-level AUC and AP (average accuracy). Experimental results show that the proposed method achieves excellent performance of 99.9%/98.7%/84.1%, 99.1%/95.3%/76.1%, 100%/98.1%/86.7% and 73.6%/69.1%/36.0% on three types of metrics on four datasets, respectively, which is better than the current advanced model.