A Deep Metric Neural Network with Disentangled Representation for Detecting Smartphone Glass Defects

• Go, Gwang-Myong ^{[2] [1]} ; Seok-Jun Bu ^[1] ; Sung-Bae Cho ^[1]
  1. [1] Yonsei University

     Yonsei University

     Corea del Sur

     
  2. [2] Samsung Electronics, Co., Ltd. Suwon 16706, South Korea
• Localización: Intelligent Data Engineering and Automated Learning – IDEAL 2020. 21st International Conference: Guimarães, Portugal; November 4–6, 2020. Proceedings / Cesar Analide (ed. lit.), Paulo Novais (ed. lit.), David Camacho Fernández (ed. lit.), Hujun Yin (ed. lit.), Vol. 2, 2020 (Part II), ISBN 978-3-030-62365-4, págs. 485-494
• Idioma: inglés
• Enlaces
  • Texto completo
• Resumen

  • For defect inspection using computer vision, deep learning models have been introduced to improve the conventional rule-based pattern analysis. A lot of data is a prerequisite to the success of them, but the on-the-spot industrial field suffers from lack of data. In this paper, we propose a deep metric neural network to improve the performance even with insufficient data imbalanced in class. The model is verified with the dataset of new products by evaluating the accuracy with 10-fold cross-validation. Our model is based on the data in the smallest category, 1.2 K, which achieves the highest performance of 90.42% using sampled pairs without using all the data for training. High accuracy has been achieved and proven applicability in the industry compared to the conventional machine learning models.