STUDY ON EXPANSION OF TRAINING DATA FOR IMPROVING CORROSION DETECTION ACCURACY OF STEEL GIRDER BRIDGES USING A CONVOLUTIONAL NEURAL NETWORK

Abstract

The number of bridges to be inspected is increasing due to aging bridges, and the number of inspection engineers is decreasing due to the declining birthrate and aging population. In recent years, the application of convolutional neural networks as the machine learning has highly effective inspection method in the civil engineering. Our previous study was developed the learning model as a corrosion detector for steel girder bridges at each area of construction and civil engineering offices using photos taken from road bridge inspection results in Fukushima Prefecture as training data, and the model was reported to have practical classification accuracy. This study investigated the possibility of improving the classification accuracy by the learning models which were built from training dataset integrated the two-areas from our previous studies. As a result, we found that the training data for corrosion increased about twice by integrating the twoareas, and the percentage of corrosion labels in the total training data was about 50%, which contributed to the improvement of approximately six points in the classification accuracy of corrosion. By combining the training data from the Aizu and/or the Kenhoku areas, the high classification accuracy for corrosion detection was achieved.