F.Blanchard, E.Bousser, J. E. Klemberg-Sapieha, L. Martinu

Department of Engineering Physics
Polytechnique Montreal
Montreal, Quebec H3T 1J4, Canada

Image segmentation is a key component of microstructural analysis when knowledge of the internal microstructure of a coating is desired. Unfortunately, segmentation is time consuming and prone to subjectivity issues when done manually. These problems are especially important when analysing coatings with chaotic pore distribution, such as thermal sprayed coatings, where analysis of multiple images for one sample is required to ensure a complete picture. Segmentation algorithms based machine and deep learning can both reduce the workload and reduce variations between samples due to human input. By training a model thoroughly once and then applying it to all the samples in a study, we can evaluate each image with the same criteria. This method also enables the identification and segmentation of an arbitrarily large number of different microstructural features, beyond simply distinguishing the pore space from the matrix, which would compound workload and judgment errors if made manually.