Deep learning corrosion detection with confidence

Corrosion costs an estimated 3–4% of GDP for most nations each year, leading to significant loss of assets. Research regarding automatic corrosion detection is ongoing, with recent progress leveraging advances in deep learning. Studies are hindered however, by the lack of a publicly available dataset. Thus, corrosion detection models use locally produced datasets suitable for the immediate conditions, but are unable to produce generalized models for corrosion detection. The corrosion detection model algorithms will output a considerable number of false positives and false negatives when challenged in the field. In this paper, we present a deep learning corrosion detector that performs pixel-level segmentation of corrosion. Moreover, three Bayesian variants are presented that provide uncertainty estimates depicting the confidence levels at each pixel, to better inform decision makers. Experiments were performed on a freshly collected dataset consisting of 225 images, discussed and validated herein.

Will Nash, Liang Zheng, Nick Birbilis