Appearance evaluation of rust on weathering steel by multiple CNN discriminator using multiple patch images

Abstract

The purpose of this study is to determine the grade of a rust image using Convolutional Neural Network (CNN). There is a problem in judging rust images using CNN that there is a variability in evaluation depending on the position from which the image is cropped. In order to improve classifica- tion accuracy, we proposes methods using multiple CNN discriminators for this judgment and a method integrating the judgment results of multiple patch images obtained from a single image by majority vote.