|About this Abstract
||Materials Science & Technology 2019
||Applications of Modern Characterization Techniques to Ferrous Alloys and Steel Products
||Applications of Machine Learning In Microscopy for Routine Assessment and Quality Assurance in Ferrous Alloy Production, Manufacturing and Service
||Roger Barnett, Alisa Stratulat
|On-Site Speaker (Planned)
All quantitative micrograph analyses involve some segmentation - dividing images into regions. E.g. segmenting grain from grain boundary to then determine grain size. This typically involves intensity thresholding with binary operation post-processing but may be difficult or time consuming.
We demonstrate a new approach for routine analysis using machine learning. A user labels small regions and then a feature vector is created for each pixel, using Gaussian, Sobel, Gabor, mean and Hessian filters plus intensity values. Using these feature vectors a “forest of random decision trees” approach is used to create a classifier which best recovers the training labels. This classifier is applied to segment the entire image, making subsequent quantitative analysis much faster and more straightforward.
We present several industrial examples using this approach for faster routine analysis of metal micrographs including automated recognition of phases, grain boundaries, layers and phases in different steel and ferrous alloys.
||Definite: At-meeting proceedings