Accelerating Corrosion Surface-Area Measurements with Computer Vision and Deep Learning: An Ensemble Approach

Abstract

The corrosion of infrastructure and facilities poses a significant challenge for the United States Department of Defense (DoD) in terms of cost and military readiness. To tackle this challenge, our research introduces a data-driven corrosion segmentation method that combines three deep learning-based models and an ensemble learning approach for the automatic identification and segmentation of corroded regions within high-resolution images. The method involves several stages, such as data annotation, preprocessing, augmentation, model implementation, and performance evaluation. The deep learning models used include Feature Pyramid Network (FPN) with Residual Network (ResNet)-34 encoder, UNet with ResNet-34 encoder, and UNet++ with Visual Geometry Group (VGG)-19 encoder. Ensemble learning, a technique that integrates these deep learning models, was employed to improve prediction accuracy and overall performance. The proposed method is evaluated using both the Dice score and the Intersection over Union (IoU) score metrics. Experimental results demonstrate that the ensemble learning approach outperforms individual models, achieving a Dice score of 90.1% and an IoU score of 83.9%. The approach shows promise to automatically detect and measure corrosion, which can reduce inspection costs and identify major issues to aid in prevention of structural failure. The tool developed in this study will be expanded to provide similar capabilities for large-scale civil infrastructure.