The aim of this investigation is to analyze the performance of several supervised machine learning algorithms for solving the automatic classification problem of steel image microstructures. We conducted an experiment using a public-domain dataset of Ultra High Carbon Steel Micrographs (UHCSM). This image database consists of a collection of scanning electron micrographs (SEM) taken from samples of a commercial roll-mill casting with a nominal carbon of 2%. Heat treatments such as annealing, water quenching, air and furnace cooling were performed on steel samples so primary microconstituents could be found in micrographs. Each of these microconstituents defines each of the categories of classification to be accomplished by machine learning algorithms. The heat treatments brought about 4 usable classes (sets of images) of primary microconstituents: pearlite, spheroidite, proeutectoid cementite network, pearlite containing spheroidite. All labeled images are prepared to improve models' accuracy in a preprocessing stage so that the image dataset is ready for feature extraction. In order to develop classification models, we put to the test distinct machine learning approaches by working with Matlab's classification learner application where we perform automated training to search for the best classification model type, including Decision Trees, Support Vector Machines (SVM), Discriminant Analysis, Nearest Neighbors, Naive Bayes, Ensemble k-NN, and Neural Network classification. For obtaining the features of the images (feature extraction) we choose the method of Bag-of-features with 400 words for the first experiment, and 327 words by removing less important features for a second experiment. The experimented models reached very different accuracy values on training, with SVM as the best classifier which gets 91.6% accuracy. We can conclude that classic machine learning algorithms solve the classification, but an accuracy improvement can be reached by investigating deep learning techniques.
Read full abstract