Proceedings of the 13th EAI International Conference on Mobile Multimedia Communications, Mobimedia 2020, 27-28 August 2020, Cyberspace

Research Article

Transfer Learning Based Screen Defect Classification

Download562 downloads
  • @INPROCEEDINGS{10.4108/eai.27-8-2020.2294670,
        author={Li  Yilei and Li  Chengyuan and Zhang  Yifan and Chang  Shuo and Zhang  Fan and wang  zixuan},
        title={Transfer Learning Based Screen Defect Classification},
        proceedings={Proceedings of the 13th EAI International Conference on Mobile Multimedia Communications, Mobimedia 2020, 27-28 August 2020, Cyberspace},
        publisher={EAI},
        proceedings_a={MOBIMEDIA},
        year={2020},
        month={11},
        keywords={defect classification transfer learning cnn generation module},
        doi={10.4108/eai.27-8-2020.2294670}
    }
    
  • Li Yilei
    Li Chengyuan
    Zhang Yifan
    Chang Shuo
    Zhang Fan
    wang zixuan
    Year: 2020
    Transfer Learning Based Screen Defect Classification
    MOBIMEDIA
    EAI
    DOI: 10.4108/eai.27-8-2020.2294670
Li Yilei1,*, Li Chengyuan1, Zhang Yifan1, Chang Shuo1, Zhang Fan2, wang zixuan3
  • 1: Laboratory of Universal Wireless Communications, Ministry of Education Beijing University of Posts and Telecommunications, Beijing, P.R.China, 100876.
  • 2: School of Information and Communication Engineering, Beijing Information Science and Technology University
  • 3: Laboratory of Universal Wireless Communications, Ministry of Education, Beijing University of Posts and Telecommunication, Beijing, P.R.China, 100876.
*Contact email: 962641150@qq.com

Abstract

For the screen defect classification, human inspectors and traditional machine learning algorithms are inefficient and inaccurate. Convolutional neural network (CNN) driven by data are feasible solutions. However, real training images are limited in the industrial scenario, which causes overfitting. Hence, in this paper, a novel learning based method is proposed for defect classification, which is based on the CNN. To alleviate the problem of limited images, two strategies are introduced into model learning. For the training data, a data generation module is implemented to enlarge the training dataset. For the CNN model learning, transfer learning is applied in the whole training process. To verify the proposed method, various experiments are carried out. The results indicate that our screen defect classification model achieves superior performance.