Deep CNN-based semi-supervised learning approach for identifying and segmenting corrosion in hydraulic steel and water resources infrastructure

Abstract

The United States faces significant challenges due to corrosion, with its impact on military and civilian infrastructure incurring over $20 billion in annual maintenance costs. The damage due to corrosion is profound, threatening structural safety, reducing esthetic value, and leading to costly repairs. To mitigate these effects, the Unified Facilities Criteria and Unified Facilities Guidance Specifications advise the use of protective coatings on metal surfaces. Early corrosion detection is crucial for maintaining structural integrity and minimizing maintenance costs. Recent breakthroughs in artificial intelligence and deep learning, including accurate corrosion classification, have significantly revolutionized the detection and management of corrosion. Despite these advancements, automatic corrosion segmentation in civil infrastructure remains challenging due to the scarcity of images and the labor-intensive annotation process. Moreover, existing segmentation methods are unable to manage the complexities that come with high-resolution corrosion images. This paper proposes a novel, semi-supervised, convolutional neural network-based image segmentation method for the automatic identification and segmentation of corrosion on coated steel surfaces, using both unlabeled and labeled corrosion images and leveraging the mean teacher model. The proposed novel method involves three steps: (1) utilizing high-resolution digital microscopy to capture detailed images and dividing them into manageable patches; (2) applying a semi-supervised learning approach, leveraging unlabeled corrosion images for enhanced segmentation precision; and (3) employing a smoothing module to improve the continuity of information. The proposed corrosion detection method has demonstrated promising performance with only 67% labeled data, achieving mean precision, recall, F-1 measure, and intersection over union of 90.0%, 96.2%, 92.7%, and 87.1%, respectively. Even with just 33% labeled data, the method maintains strong performance when compared to fully supervised deep learning models. This demonstrates a substantial data resource saving while ensuring accurate and reliable corrosion detection, which is crucial for infrastructure health monitoring. The successful validation of this approach provides a method that dramatically reduces the amount of visual data required to generate a reliable model.