• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2021, Vol. 57 ›› Issue (17): 86-97.doi: 10.3901/JME.2021.17.086

• 特邀专栏:智能制造前沿及应用 • 上一篇    下一篇

扫码分享

基于两阶段深度迁移学习的面料疵点检测算法

赵树煊1,2, 张洁1,2, 汪俊亮1,2, 徐楚桥3   

  1. 1. 东华大学人工智能研究院 上海 201620;
    2. 上海工业大数据与智能系统工程技术研究中心 上海 201620;
    3. 上海交通大学机械与动力工程学院 上海 200240
  • 收稿日期:2021-01-15 修回日期:2021-04-20 发布日期:2021-11-16
  • 通讯作者: 张洁(通信作者),女,1963年出生,博士,教授,博士研究生导师。主要研究方向为智能制造系统、工业大数据。E-mail:mezhangjie@dhu.edu.cn
  • 作者简介:赵树煊,男,1997年出生。主要研究方向为机器视觉与深度学习。E-mail:dhuzhaoshuxuan@163.com
  • 基金资助:
    国家自然科学基金(51905091)、上海市教育发展基金会和上海市教育委员会“晨光计划”(20CG41)和上海市科技计划(20DZ2251400)资助项目。

Fabric Defect Detection Algorithm Based on Two-stage Deep Transfer Learning

ZHAO Shuxuan1,2, ZHANG Jie1,2, WANG Junliang1,2, XU Chuqiao3   

  1. 1. Institute of Artificial Intelligence, Donghua University, Shanghai 201620;
    2. Shanghai Industrial Big Data and Intelligent Systems Engineering Technology Center, Shanghai 201620;
    3. School of Mechnical Engineering, Shanghai Jiao Tong University, Shanghai 200240
  • Received:2021-01-15 Revised:2021-04-20 Published:2021-11-16

摘要: 针对针织格纹面料疵点检测存在检测实时性差和疵点数据稀缺的问题,提出一种基于两阶段深度迁移学习的面料疵点检测算法,实现对疵点的实时高精度检测与检测模型的高效训练。第一阶段迁移:设计面料疵点先验知识迁移算法,通过聚类算法求得交并比最优的四类疵点预选框尺寸参数,使用带有先验知识的疵点预选框替代基于特征的定位方法,实现面料疵点尺寸特征先验知识的迁移,提高面料疵点的定位速度;第二阶段迁移:设计面料特征提取能力迁移算法,利用不同种面料之间具有通用特征的特性,通过将纯色棉麻布检测模型参数迁移至格纹面料检测模型,实现对面料通用特征提取能力的迁移,减少检测模型训练所需的样本数量,提高检测模型训练效率。实验结果表明,在检测性能方面,提出的面料疵点检测算法检测精度为95%、检测速度可达30 m/min,优于传统的目标检测算法,能够满足面料生产中对于检测性能的要求;在模型训练方面,检测模型训练所需疵点样本数量减少50%以上、检测模型训练速度同比提高3倍。

关键词: 面料疵点检测, 深度学习, 迁移学习

Abstract: To solve the problem of low real-time detection and lack of defect images data in fabric defect detection. The fabric defect detection algorithm based on two-stage deep transfer learning is proposed, realizing the real-time detection and efficient training of four types of defects in plaid fabric. First stage transfer, fabric defect prior knowledge transfer algorithm which takes the feature of defects as the transfer object is designed and four kinds of defect pre-boxes parameters with optimal intersection ratios are obtained by clustering algorithm. The pre-boxes with prior knowledge are used to replace the method based on fabric features in defects location, realizing the transfer of the fabric defect features' prior knowledge and improving the location speed of fabric defects. Second stage transfer, a fabric feature extraction ability transfer algorithm is designed to take advantage of the universal feature among different fabrics, the universal feature extraction abilities of convolutional neural network are realized by transfering parameters of linen detection model to patterned plaid detection model, so as to reduce the number of samples that model training need and improve the effectiveness of model training. The experimental results show that, in terms of detection performance, the detection algorithm we proposed has a detection accuracy of 95%, which is better than traditional algorithm and can satisfy the requirements of detection. In terms of model training, the number of fabric defects are reduced at least 50%, and the speed of model training is increased by more than 3 times.

Key words: farbic defects detection, deep learning, transfer learning

中图分类号: