Microstructure Classification of Ultra High Carbon Steel Using Deep Learning Approach

One way of understanding the material characteristics is using the information obtained from the micrographs of the materials. Classification and evaluation of microstructures is done by human experts leading to errors. Computer vision approaches are superior in this scenario. In this work, a deep learning approach is presented for classifying the Ultra High Carbon Steel (UHCS) micrographs obtained from SEM (Scanning Electron Microscope) images captured under variety of heat treatments. There are seven classes and 961 images in the UHCS dataset. In this proposed work, several (652) pre-trained Convolutional Neural Networks (CNN) are used for extracting the features from the micrographs. Selected features from the CNNs are concatenated and are given to a classifier. Feature vector size is reduced by using Boruta feature selection algorithm. Since the dataset is an imbalanced one, oversampling strategy ADASYN is employed for the under-represented classes like Martensite, and Pearlite + Widmanstatten categories. A balanced accuracy of 95.2% and F1 score of 97% is obtained for a 10-fold cross validation using a Passive Aggressive Classifier. Further, the proposed approach is also tested with another dataset using Ti–6Al–4V alloy micrographs. There are three classes (Acicular(186), Bimodal(350) and Lamellar(689) and 1225 images in this dataset. A classification accuracy of well above 98% is obtained on this dataset.