| Abstract Scope |
In quantitative microscopy, microstructure image segmentation is essential for image analysis and materials characterization. The rising deep convolutional neural network methods for semantic segmentation in natural images have recently transferred to materials images and demonstrated outstanding performance in complicated microstructure datasets. However, current supervised learning solutions require pixel-level human annotations, which is painstaking, biased and infeasible for some cases. It is worth exploring the possibility to go from fully supervised to semi-supervised and unsupervised methods with the goal of achieving comparable performance while alleviating the annotation cost. Here, we demonstrate our effort in moving from supervised methods to unsupervised methods. Multiple aspects of the strategies including dataset generation, annotation, transfer learning, evaluation, etc., are discussed. |