Abstract
Corrosion detection in industrial assets and components is an important broad problem in the industries since it allows the temporal tracking of possible issues and the execution of preventive maintenance actions, such as protective coating. However, solving this problem using modern machine learning methods usually demands a careful design of artificial intelligence tools, such as neural networks, high computational resources for training and inference, and a large and adequate dataset. In this work we investigate the application of deep convolutional neural networks to the problem of image semantic segmentation of superficial corrosion and dirt present in mining industrial assets, using a set of images collected in place by corrosion inspectors and manually labeled by a data team. We compare two networks based on the popular U-Net model, in which one of them uses the transferred features from a pre-trained VGG-16 image classification model. Our main contribution is to provide insights about the application of deep neural networks in this particular domain, mostly regarding the size and quality of the constructed dataset, the existing computational resources constraints, and the observed benefits of using a pre-trained model, as well as discuss some preliminary segmentation results obtained. Our main results show the practical usefulness of transfer learning approaches, which presents significantly better results in our dataset, and also highlight the importance of constructing a dataset with appropriate size and sample quality.
