Deep learning AI for corrosion detection

Will Nash, Tom Drummond, Nick Birbilis

Research output: Chapter in Book/Report/Conference proceedingConference PaperOther

Abstract

Visual inspection is a vital component of asset management that stands to benefit from automation. Using artificial intelligence to assist inspections can increase safety, reduce access costs, provide objective classification, and integrate with digital asset management systems. The work presented herein investigates the impact of dataset size on Deep Learning for automatic detection of corrosion on steel assets. Dataset creation is typically one of the first steps when applying Machine Learning methods to a new task; and the real-world performance of models hinges on the quality and quantity of data available. Producing an image dataset for semantic segmentation is resource intensive, particularly for specialist subjects where class segmentation is not able to be effectively farmed out. The benefit of producing a large, but poorly labelled, dataset versus a small, expertly segmented dataset for semantic segmentation is an open question. Here we show that a large, noisy dataset outperforms a small, expertly segmented dataset for training a Fully Convolutional Network model for semantic segmentation of corrosion in images. A large dataset of 250 images with segmentations labelled by undergraduates and a second dataset of just 10 images, with segmentations labelled by subject matter experts were produced. The mean Intersection over Union and micro F-score metrics were compared after training for 50,000 epochs. The relationship between dataset size and F-score was investigated to estimate the requirements to achieve human level accuracy. This work is illustrative for researchers setting out to develop deep learning models for detection and location of specialist features.

Original languageEnglish
Title of host publicationCorrosion 2019
PublisherNACE International
Number of pages11
Publication statusPublished - 2019
EventNACE International - Corrosion 2019 - Nashville, United States of America
Duration: 24 Mar 201928 Mar 2019

Publication series

NameNACE - International Corrosion Conference Series
ISSN (Print)0361-4409

Conference

ConferenceNACE International - Corrosion 2019
CountryUnited States of America
CityNashville
Period24/03/1928/03/19

Keywords

  • Corrosion
  • Datasets
  • Fully Convolutional Network
  • Machine Learning
  • Semantic Segmentation

Cite this

Nash, W., Drummond, T., & Birbilis, N. (2019). Deep learning AI for corrosion detection. In Corrosion 2019 (NACE - International Corrosion Conference Series). NACE International.
Nash, Will ; Drummond, Tom ; Birbilis, Nick. / Deep learning AI for corrosion detection. Corrosion 2019. NACE International, 2019. (NACE - International Corrosion Conference Series).
@inproceedings{4cd8086c80ea425484504d3833ddea32,
title = "Deep learning AI for corrosion detection",
abstract = "Visual inspection is a vital component of asset management that stands to benefit from automation. Using artificial intelligence to assist inspections can increase safety, reduce access costs, provide objective classification, and integrate with digital asset management systems. The work presented herein investigates the impact of dataset size on Deep Learning for automatic detection of corrosion on steel assets. Dataset creation is typically one of the first steps when applying Machine Learning methods to a new task; and the real-world performance of models hinges on the quality and quantity of data available. Producing an image dataset for semantic segmentation is resource intensive, particularly for specialist subjects where class segmentation is not able to be effectively farmed out. The benefit of producing a large, but poorly labelled, dataset versus a small, expertly segmented dataset for semantic segmentation is an open question. Here we show that a large, noisy dataset outperforms a small, expertly segmented dataset for training a Fully Convolutional Network model for semantic segmentation of corrosion in images. A large dataset of 250 images with segmentations labelled by undergraduates and a second dataset of just 10 images, with segmentations labelled by subject matter experts were produced. The mean Intersection over Union and micro F-score metrics were compared after training for 50,000 epochs. The relationship between dataset size and F-score was investigated to estimate the requirements to achieve human level accuracy. This work is illustrative for researchers setting out to develop deep learning models for detection and location of specialist features.",
keywords = "Corrosion, Datasets, Fully Convolutional Network, Machine Learning, Semantic Segmentation",
author = "Will Nash and Tom Drummond and Nick Birbilis",
year = "2019",
language = "English",
series = "NACE - International Corrosion Conference Series",
publisher = "NACE International",
booktitle = "Corrosion 2019",

}

Nash, W, Drummond, T & Birbilis, N 2019, Deep learning AI for corrosion detection. in Corrosion 2019. NACE - International Corrosion Conference Series, NACE International, NACE International - Corrosion 2019, Nashville, United States of America, 24/03/19.

Deep learning AI for corrosion detection. / Nash, Will; Drummond, Tom; Birbilis, Nick.

Corrosion 2019. NACE International, 2019. (NACE - International Corrosion Conference Series).

Research output: Chapter in Book/Report/Conference proceedingConference PaperOther

TY - GEN

T1 - Deep learning AI for corrosion detection

AU - Nash, Will

AU - Drummond, Tom

AU - Birbilis, Nick

PY - 2019

Y1 - 2019

N2 - Visual inspection is a vital component of asset management that stands to benefit from automation. Using artificial intelligence to assist inspections can increase safety, reduce access costs, provide objective classification, and integrate with digital asset management systems. The work presented herein investigates the impact of dataset size on Deep Learning for automatic detection of corrosion on steel assets. Dataset creation is typically one of the first steps when applying Machine Learning methods to a new task; and the real-world performance of models hinges on the quality and quantity of data available. Producing an image dataset for semantic segmentation is resource intensive, particularly for specialist subjects where class segmentation is not able to be effectively farmed out. The benefit of producing a large, but poorly labelled, dataset versus a small, expertly segmented dataset for semantic segmentation is an open question. Here we show that a large, noisy dataset outperforms a small, expertly segmented dataset for training a Fully Convolutional Network model for semantic segmentation of corrosion in images. A large dataset of 250 images with segmentations labelled by undergraduates and a second dataset of just 10 images, with segmentations labelled by subject matter experts were produced. The mean Intersection over Union and micro F-score metrics were compared after training for 50,000 epochs. The relationship between dataset size and F-score was investigated to estimate the requirements to achieve human level accuracy. This work is illustrative for researchers setting out to develop deep learning models for detection and location of specialist features.

AB - Visual inspection is a vital component of asset management that stands to benefit from automation. Using artificial intelligence to assist inspections can increase safety, reduce access costs, provide objective classification, and integrate with digital asset management systems. The work presented herein investigates the impact of dataset size on Deep Learning for automatic detection of corrosion on steel assets. Dataset creation is typically one of the first steps when applying Machine Learning methods to a new task; and the real-world performance of models hinges on the quality and quantity of data available. Producing an image dataset for semantic segmentation is resource intensive, particularly for specialist subjects where class segmentation is not able to be effectively farmed out. The benefit of producing a large, but poorly labelled, dataset versus a small, expertly segmented dataset for semantic segmentation is an open question. Here we show that a large, noisy dataset outperforms a small, expertly segmented dataset for training a Fully Convolutional Network model for semantic segmentation of corrosion in images. A large dataset of 250 images with segmentations labelled by undergraduates and a second dataset of just 10 images, with segmentations labelled by subject matter experts were produced. The mean Intersection over Union and micro F-score metrics were compared after training for 50,000 epochs. The relationship between dataset size and F-score was investigated to estimate the requirements to achieve human level accuracy. This work is illustrative for researchers setting out to develop deep learning models for detection and location of specialist features.

KW - Corrosion

KW - Datasets

KW - Fully Convolutional Network

KW - Machine Learning

KW - Semantic Segmentation

UR - http://www.scopus.com/inward/record.url?scp=85070075966&partnerID=8YFLogxK

M3 - Conference Paper

AN - SCOPUS:85070075966

T3 - NACE - International Corrosion Conference Series

BT - Corrosion 2019

PB - NACE International

ER -

Nash W, Drummond T, Birbilis N. Deep learning AI for corrosion detection. In Corrosion 2019. NACE International. 2019. (NACE - International Corrosion Conference Series).