Automated Corrosion Detection Using Crowd Sourced Training for Deep Learning

TL;DR

This paper presents a cloud-based system that uses crowd-sourced image labeling to train deep learning models for automated corrosion detection from images and videos, enabling remote, fast, and cost-effective monitoring.

Contribution

It introduces a novel crowd-sourced labeling platform integrated with a deep learning model for corrosion detection, demonstrating its effectiveness over one month.

Findings

Successful crowd-sourced labeling improved model training.

The system enabled rapid corrosion detection in images and videos.

Cost and time savings in corrosion monitoring were achieved.

Abstract

The automated detection of corrosion from images (i.e., photographs) or video (i.e., drone footage) presents significant advantages in terms of corrosion monitoring. Such advantages include access to remote locations, mitigation of risk to inspectors, cost savings and monitoring speed. The automated detection of corrosion requires deep learning to approach human level artificial intelligence (A.I.). The training of a deep learning model requires intensive image labelling, and in order to generate a large database of labelled images, crowd sourced labelling via a dedicated website was sought. The website (corrosiondetector.com) permits any user to label images, with such labelling then contributing to the training of a cloud based A.I. model - with such a cloud-based model then capable of assessing any fresh (or uploaded) image for the presence of corrosion. In other words, the website includes both the crowd sourced training process, but also the end use of the evolving model. Herein, the results and findings from the website (corrosiondetector.com) over the period of approximately one month, are reported.