2022 Volume 68 Issue 5 Pages 87-97
A conventional shuttle loom was equipped with an industrial camera to capture textile images while weaving. Both defect-free textiles and four types of defect textiles were weaved for five kinds of patterns. The defect textiles were intentionally added the defects at the weaving stage. A unique convolutional autoencoder was individually trained on the normal image data of these textiles, and differences were taken between defective input and output images using a trained model. Model accuracies of defect segmentation were verified by calculating the AUC values taken from residual maps and ground truths prepared by visual inspection. The time until the model achieved satisfactory accuracy was also examined. The resulting model achieved extremely high accuracy for specific defects, regardless of patterns. Longer one repeating pattern required longer training time, although training time was short enough to train while weaving. Thus, this study showed the feasibility of an automatic learning defect segmentation system to train for each pattern individually.
