Corrosion grade recognition for weathering steel plate based on a convolutional neural network
Wang Yan, Xiaoli Shen, Kai Wu, Mingquan Huang
Abstract
Abstract For the maintenance of weathering steel structure facilities, it is necessary to evaluate the corrosion grade of the rust layer on the surface regularly. At present, corrosion grade classification of weathering steel is mainly based on visual inspection with the human eye. In this paper, a deep learning method using a convolutional neural network (CNN) to evaluate the corrosion grade of weathering steel is proposed to save time and manpower. Firstly, the image dataset of the corrosion steel plate was established using salt spray tests. Then, a CNN architecture named VGG-Corrosion was designed to evaluate the corrosion grade of the corroded steel plate. The effects of the learning rate, transfer learning, and batch size were also investigated to clarify the best hyperparameter configurations to train a powerful corrosion grade classification model. Under the best combination of considered hyperparameters, the mean average accuracy for the corrosion grade evaluation of the test results is 90.96%. The test results indicated that the CNN-based corrosion grade recognition for weathering steel plate is prospective, which would be helpful for safety evaluation of steel structures.