Enhancing the confidence of segmenetation maps using complimentary data and convolutional segmentation

点击翻译按钮获取中文摘要

Accurate maps of the distribution of phases inside multiphase materials is vital for extracting characteristic metrics (such as volume fractions and surface areas), as well as for specifying mesh domains for use in multiphysics simulations. Various high resolution imaging techniques have been developed for this task; however, each has its own strengths and weaknesses. For example, focused ion-beam scanning electing microscopy (FIB-SEM) can achieve very high resolution (c. 5 nm) and phase sensitivity, but struggles to capture representative volumes and is a destructive technique. Alternatively, X-ray computed tomography (XCT) can capture large volumes in-situ, but often performs poorly at differentiating phases, particularly of small features.By combining image data from multiple techniques (or even just multiple modes or detectors of a single technique) it is possible segment data (i.e. assign each pixel to a phase) with more confidence than any individual method. However, this requires the use of convolutional methods that are able to combine the information from various input channels. In particular, deep convolutional neural networks are very well suited to this kind of task and have seen dramatic advancements in their implementation over the last 5 years due, in part, to their use in self driving cars.This project would develop machine learning based convolutional segmentation tools for confidently characterising multimodal image data, with potential for impact across all areas of materials science.