In quantitative microscopy, image segmentation plays a central role in quantitative measurements and analyses of microstructure constituents. Developing effective and efficient methods for automating the segmentation process is highly valued for materials research and manufacturing. Conventional image processing-based segmentation methods reach their limits in handling complicated microstructures and often require sophisticated processing pipelines. The cutting-edge data-driven deep learning methods have made huge breakthroughs in image-based tasks and they provide new possibilities for advanced microstructure image segmentation.

In this thesis, we demonstrate that deep learning methods can be successfully applied to microstructure image segmentation tasks with great advantages in automation, performance and generality. We demonstrate the capabilities of deep learning methods in two different problems:​ (1) segmentation of complex multi-constituents microstructures in ultrahigh carbon steel and (2) segmentation of low-contrast lath-shaped bainite in complex phase steel. General guidelines and strategies for tackling such tasks are discussed.

To alleviate annotation cost and achieve high efficiency for practical applications, we develop our deep learning models in a semi-supervised manner with a significantly reduced amount of annotated data. An automated training image selection algorithm is proposed and we demonstrate in two steel microstructure segmentation datasets that deep learning models trained by one or a few images are competitive with fully-supervised models using 4 times more training images.