Corrosion Classification Study of Mild Steel in 3.5% NaCl using Convolutional Neural Networks

Abstract

Corrosion detection using advanced equipment could be sometimes unavailable in resource-limited settings. To make up for the corrosion testing gap, image capturing and processing with Convolutional Neural Networks (CNN) have gained prominence in corrosion studies. In this study, two CNN models were built and trained using images taken with a mobile phone camera and a digital microscope. The CNN models were built to categorize corroded images into three different classes based on the surface area of the sample that were covered by the corrosion products. The study shows that CNN corrosion classifiers perform very well with accuracy above 80% for both models. The use of CNN was found to be effective for multiclass corrosion.