2020 年 66 巻 3 号 p. 47-54
In this paper, we propose a convolutional autoencoder with a new structure for unsupervised learning when the purity of the training data is not guaranteed. This autoencoder has two unique features: the target area is reconstructed from the surrounding areas and the L2 loss is predicted simultaneously. The superiority of this model was verified using SEM images of defective nanofibrous materials by calculating the AUC value. The results of our experiments with the training data contaminated by defective data show that the former feature improves the robustness against contamination of the training data and the latter improves the accuracy. Although this approach did not achieve the highest accuracy, it could reduce the cost of annotation for practical use. Furthermore, we applied our method to images of NISHIJIN textiles and found that it worked well for some types of textiles.