Corrosion damage detection and evaluation of coated steel components under multiple illumination conditions

Abstract

This study proposes a corrosion grade prediction network for coated steel components under varying illumination conditions. Images of corroded steel components were captured under low illumination, ambient, and high illumination conditions by adjusting camera parameters in a field workshop. Three parallel enhanced Mobile-Vision-Transformer networks were developed to assess prediction performance for corrosion grades under different illumination conditions and two transfer learning approaches. Network weights were fused, incorporating a global average pooling layer and convolution layer to enable direct corrosion grade prediction across varied illumination conditions. The impact of learning rate, input image size, image augmentation technique, etc., on network performance was investigated. The interpretability of the network is enhanced using the gradient-weighted class activation mapping method. Furthermore, prediction accuracy was verified using images of corroded coated steel plates captured under diverse illumination conditions and corrosion grades from accelerated laboratory corrosion tests. Finally, a graphical user interface was designed for automated corrosion grade prediction in coated steel components under varying illumination conditions.