Visual inspection and characterization of external corrosion in pipelines using deep neural network. (October 2019)

• Record Type: Journal Article
• Authors: Bastian, Blossom Treesa; N, Jaspreeth; Ranjith, S. Kumar; Jiji, C.V.
• Is Part Of: NDT & E international. Volume 107(2019)
• Publication Date: 2019-10
• Subjects: Water/oil pipelines -- Corrosion detection -- Optical inspection -- Deep learning -- Convolutional neural networks; Nondestructive testing -- Periodicals; Contrôle non destructif -- Périodiques; Electronic journals; 620.1127
• DOI: 10.1016/j.ndteint.2019.102134
• Languages: English
• ISSNs: 0963-8695

Abstract

Abstract: In this paper, we proposed a computer vision based approach to detect corrosion in water, oil and gas pipelines. For this, we created a dataset containing more than 140, 000 optical images of pipelines with different levels of corrosion. A custom designed convolutional neural network (CNN) was applied to classify the images of pipelines based on their corrosion level. This in-house fabricated CNN has very few parameters to be learned in comparison with the existing CNN classifiers. However, it produced significantly higher classification accuracy (98.8%) with an ability to discriminate between images of corroded pipelines and images without corrosion but having patterns similar to corroded pipelines. The proposed network surpassed most of the state-of-the-art classifiers in its performance. In addition, we proposed a localisation algorithm based on a recursive region-based method, to selectively identify the corroded regions in a given image with higher precision. The proposed deep learning approach effectively wards off the need for manual inspection and other non-vision based non-destructive evaluation techniques for pipeline corrosion which are cost ineffective and interrupts the functioning of pipelines.