Thinning Evaluation of Steel Plates for Weathering Tests Based on Convolutional Neural Networks

Abstract

Weathering tests using monitored steel plates are a widely applied method for evaluating the atmospheric corrosion rate in Japan. To calculate the regional corrosion rate, the corrosion layer on the surface of the steel plate needs to be removed to determine the thinning. However, the process of removing the corrosion layer is time and labor consuming. To tackle this issue, this study proposed an image recognition method based on convolutional neural networks (CNNs) to evaluate the thinning of weathering test samples. To this end, the existing data collected from the weathering tests were reused to generate a dataset named “Corrosion-Fukui” that consisted of 77 raw images labeled with their numerical extent of thinning. To generate more samples for training, a criteria based on thinning extent that classified the raw images into six corrosion levels were defined to implement cropping operation on the raw images with uniform corrosion morphology. Correspondingly, the raw images of the corroded samples with uniform corrosion morphology were chosen as “training” and “validation samples” to be cropped into small pieces labeled with the corrosion levels, whereas other raw images with nonuniform corrosion morphology were chosen as “test samples.” The performance of the proposed baseline model VGGGAP as well as three state-of-art CNN models was cross-validated on the augmented dataset and tested upon the test images using a sliding window method. The evaluation results of the 17 testing samples indicated that the corrosion thinning of the weathering test samples can be directly evaluated more efficiently from digital images using CNNs than using conventional corrosion removal methods.