Classification and segmentation (pixel-wise assignment of material phase) of material structures from microscopy images are key steps of quantitative microstructural analysis to capture the complex process-structure-property dependencies for manufacturing process optimization or quality control. Traditional microstructure quantification requires numerous manual measurements [1], is time consuming and prone to bias. Even automated microscopy image segmentation using classical computer vision methods is not always sufficient due to implementation difficulties and lack of robustness [2]. Recently, pretrained convolutional neuronal networks (CNN) have produced superior microscopy classification and segmentation results [3,4,5,6,7] indicating significant potential to perform efficient, robust and objective microstructure analysis while dramatically reducing operator input, cycle time and cost.

The scope of this work is to setup a microstructure segmentation method using the fast.ai library [8], test functionality for optical and SEM images from manufacturing environments and deploy it as part of an automated quantitative microstructure analysis procedure. Semantic segmentation (binary and multi-label) is performed using CNNs with encoder-decoder type architectures, offering state-of-the-art performance on benchmark datasets. The presented results are generated with a Resnet34 as encoder (pretrained on ImageNet) and the Unet decoder architecture. The segmentation accuracy is assessed via accuracy metrics/maps and quantitative microstructural information is extracted via an automated segmentation/analysis procedure (i.e. grain size/shape distribution). 

References

[1] ASTM E112-10, 2010, 96, 1-27.

[2] J. Stuckner Computational Materials Science, 2017, 139, 320-329.

[3] B.L. DeCost Microscopy and Microanalysis, 2019, 25, 21-29.

[4] A. Goetz Computational Materials, 2022, 8, 1-13.

[5] N.M. Senanayake Integrating Materials Manufacturing Innovation, 2020, 9, 446-458.

[6] R. Cohn JOM, 2021, 73, 2159-2172.

[7] S. Akers Computational Materials, 2021, 7, 1-9.

[8] https://www.fast.ai/