A comparison of deep learning techniques for corrosion detection

Bolton, Tom; Bass, Julian; Gaber, Tarek

Abstract

Corrosion - degradation in metal structures - is problematic, expensive to rectify, and can be unpredictable in the rate at which it spreads. Traditional preventative maintenance techniques are complemented by human visual inspection, in turn complemented by artificial intelligence vision techniques. The primary objective of this paper was to determine the most accurate deep learning model for use in corrosion detection; to achieve this, we devised an experimental comparison that tested five machine learning algorithms for the detection of corrosion from image data. The deep learning that forms the basis of algorithms used to solve object recognition problems traditionally requires large amounts of training data. As this data requires manual labelling by a person who is expert in the domain of corrosion, it is difficult and expensive
to obtain; time and expense that increase considerably as more sophisticated pixel-level annotation is applied. We discovered that high levels of accuracy (98%) can be achieved using deep learning to detect corrosion using samples annotated with simple, image-level labels. We achieved this headline accuracy through the application of transfer learning using models that had been trained on the ImageNet dataset. With many deep learning algorithms to choose from, we systematically determined the most accurate model to use as a basis for further experimentation.

• Presentation Conference Type: Conference Paper (published)
• Conference Name: Advanced Intelligent Systems and Informatics 2022
• Start Date: Nov 20, 2022
• End Date: Nov 22, 2022
• Acceptance Date: Jun 2, 2022
• Online Publication Date: Nov 17, 2022
• Publication Date: Nov 18, 2022
• Deposit Date: Aug 22, 2022
• Publicly Available Date: Nov 18, 2023
• Journal: Proceedings of the 8th International Conference on Advanced Intelligent Systems and Informatics 2022
• Publisher: Springer
• Pages: 189-198
• Series Title: Lecture Notes on Data Engineering and Communications Technologies
• Series ISSN: 2367-4512
• Edition: 1
• ISBN: 9783031206009
• DOI: https://doi.org/10.1007/978-3-031-20601-6_18
• Publisher URL: https://doi.org/10.1007/978-3-031-20601-6_18
• Additional Information: Event Type : Conference