Indexing steel corrosion in ten levels through convolutional neural network models

Abstract

Corrosion poses significant challenges to steel structures, making early detection essential to avoid high maintenance costs and potential disasters. Automated detection and extent classification enable precise damage assessment, aiding targeted repair strategies and efficient resource allocation. Traditionally, human surveyors have conducted inspections, but these manual methods are labour-intensive and time-consuming and may yield inconsistent results. To overcome the conventional limitations, this study proposes applying image processing and a convolutional neural network (CNN) to corrosion detection, using a database of 9920 images captured with regular, portable cell phone cameras. The method uses 15 well-known pre-trained models (AlexNet, DenseNet-121, EfficientNet-B0, GoogleNet, Inception-v3, MobileNet-v1, NASNet, ResNet-18, ResNet-50, ResNet-101, ShuffleNet-v1, SqueezeNet, VGG16, VGG19 and Xception). CNN models are used to detect corrosion at ten levels, based on the percentage extent of corrosion, with levels ranging from 0% to 10% (level 1) and from 90% to 100% (level 10). After comparing 15 pre-trained models, Xception was chosen as the best, achieving a mean accuracy of 0.72, an accuracy of 0.93 and F1 and recall values of 0.91 and 0.94, respectively. This automated approach can aid early detection of corrosion, enhance maintenance prioritisation and reduce inspection costs for steel structures.