Deep Residual Network-Based Automated Fabric Inspection
Chao-Ching Ho1, Wei-Chi Chou1, Bing-Han Zhou1, Eugene Su1, Yu-Zuo Liao2, Wen-Zheng Xu2 and Kuei-Chi Lee2
1Graduate Institute of Manufacturing Technology, National Taipei University of Technology, Taiwan; 2Taiwan Textile Research Institute, Taiwan
Abstract: In response to the rapid development of smart manufacturing, the use of artificial intelligence to solve industrial problems is increasing. This Research uses the supervised deep convolutional neural network to learn the characteristics of holes, yarn leakage and neural breaks in textiles, and uses the transfer learning and fine-tuning to improve training speed and accuracy. After the trained neural network achieved the goal of automated detection, the neural network used in this Research takes 0.189 seconds to predict a 600 × 600 image under the hardware architecture of 2080Ti.
Keywords: Deep convolutional neural networks, Defect identification, Fabric defect, Automated optical inspection, Digital image processing.