基于FSFF-UCDAE的无监督织物疵点检测

打开文本图片集

关键词织物疵点检测；去噪自编码器；无监督学习；全尺度特征融合；图像重构

中图分类号：TS104.3；TP183 文献标志码：A

DOI：10.13338/j.issn.1006-8341.2025.03.006

Unsupervised fabric defect detection based on FSFF-UCDAE

ZHANG Zhouqiang 1，2 ， WANG Kangxu 1 ，LI Cheng¹，ZHANG Jinxu 1 ，WANG Lin 1

（1. School of Electronics and Information，Xi'an Polytechnic University，Xi'an 710o48，China; 2.Shaanxi KeyLaboratory of Functional Apparel Fabrics，Xi'an Polytechnic University，Xi'an 71o048，China）

AbstractSupervised models heavily rely on scarce defect samples and require extensive manual annotation，making it challenging to meet practical application needs. To address this issue，the research team proposed an unsupervised fabric defect detection method based on image reconstruction and residual processing. During the training phase，only defect-free samples were used as the training set. By constructing and training a reconstruction model，the method learns and represents the normal structural characteristics of fabrics. In the detection phase，the model reconstructs the images to be inspected and calculates the residuals between the original and reconstructed images.

Thresholding and morphological processing was then applied to accurately extract and locate fabric defect regions. Experimental results demonstrate that the method not only effectively avoids reliance on defect samples but also accurately detects fabric defect regions. Its high accuracy and robustness are further ralidate by comparing with other models.

Keywordsfabric defect detection; autoencoder;unsupervised learning; full-scale feature fusion; image reconstruction

0 引言

纺织业是我国国民经济的重要支柱产业，在国家经济快速发展过程中占有重要地位[1]。（剩余13911字）