CORROSION DETECTION FOR STEEL GIRDER BRIDGES APPLY CONTRAST ADJUSTMENT AS IMAGE PREPROCESSING USING A CONVOLUTIONAL NEURAL NETWORK

Abstract

There are concerns about the infrastructure aging that was built in large numbers during the period of rapid economic growth in Japan, as well as damage to bridges caused by natural disasters. In recent years, a bridge damage detection using convolutional neural networks as a machine learning has been conducted, but efficient preparation for training data is still an issue. This study applied four types of preprocessing (flip horizontal, contrast enhancement, contrast reduction, and histogram flattening) to training images which were extracted from the results of road bridge inspections conducted by Fukushima Prefecture, furthermore, unadjusted and preprocessed training images were integrated to expand training dataset. As a result, we found that the learning model developed by integrating the unadjusted training images with the contrast-reduced training images which is to an increase of 1.5 times in the training images, increased the corrosion classification accuracy by 7%.