Vision-Based Corrosion Identification Using Data-Driven Semantic Segmentation Techniques

Corrosion is a natural process that degrades metal-made materials. Its detection is of primordial importance for quality control and for ensuring longevity of metal-made objects
in various contexts, in particular in industrial environments. Different techniques for
corrosion identification including ultrasonic testing, radio-graphic testing, and magnetic flux leakage have been proposed in the past. However, these require the use of costly
and heavy equipment onsite for successful data acquisition. An under-explored alternative is to deploy conventional lightweight and inexpensive camera systems and computer vision based methods to tackle the former problem. In this work we present a detailed benchmark of four state-of-the-art supervised semantic segmentation techniques, for vision-based pixel-level corrosion identification. We focus our study on four, recently proposed deep learning architectures which have surpassed human-level accuracy on various visual tasks. The results demonstrate that the former approaches may be used for the problem of segmenting highly irregular patterns in industrial settings, such as corrosion, with high accuracy rates.