Fast segmentation of 3D and 4D tomography using deep learning

Kurzfassung

The continuously increasing brilliance of synchrotron sources as well as the use of fast imaging detectors is able to give access to a vast amount of three- or four dimensional data to material scientists. Human-based image segmentation of complex multi-phase microstructures can easily take 100s hours of operating time and may act as a bottleneck during the research process. The development of machine learning tools and especially convolutional neural networks (CNNs) has recently shown a high impact on image segmentation tasks with a very broad range of applications, including materials science. However, the segmentation of materials’ microstructure has its very specific challenges. For that reason, the implementation of a deep CNN using a pixel-wise weighted error function is presented. The function takes into account microstructural features that are rather difficult to identify or play a crucial role for the correct description of the investigated microstructures. The benefit of the application of the trained CNN is presented on the basis of synchrotron tomography of two different alloys. Firstly, the results show that the use of a CNN is able to reduce the operation time for the segmentation of the complex microstructure of cast Al-Si alloys to <1% of the time needed with human-based segmentation. Moreover, the fully automatic segmentation increases the objectivity of a segmentation compared to a human-based one. Secondly, the application is extended to a time series of three dimensional synchrotron data for the investigated alloy to the purpose of quantifying damage evolution caused by mechanical stress.