Basic study on rust appearance evaluation of weathering steel materials using feature amount extracted by CNN and dimensionality reduction visualization method

Abstract

Currently, the general inspection method for weathering steel is to visually evaluate the rust condition in five levels, however individual differences in judgment are likely to occur. Therefore, a quantitative and automatic evaluation method using a convolutional neural network (CNN) has been proposed. However, there are problems that the rating of a small piece of rust image used as teacher data is unknown. Additionally, the CNN can only judge the rust appearance rating of a small piece of image. In this study, the feasibility of classifying the rust appearance rating was examined by visualizing the features extracted from a CNN that has not been trained on rust images and reducing the dimensionality of the features using an unsupervised learning method t-distributed probabilistic nearest neighbor embedding (t-SNE). The results showed the possibility of classifying the rust appearance rating based on the distribution of the image of small pieces of each appearance rating.