Development of a radial basis function neural network for the recognition of common phases present in carbon steel metallographs

2^nd International Conference and Exhibition on Materials Science & Engineering

October 07-09, 2013 Hampton Inn Tropicana, Las Vegas, NV, USA

Sara E. McCaslin and Adarsh Kesireddy

Scientific Tracks Abstracts: J Material Sci

Abstract :

While research has been done in applying digital image processing and neural networks to the recognition of phases and inclusions in materials such as cement and iron, little to no research has focused on recognition of steel phases. This research focuses on the automated recognition of ferrite, pearlite, cementite, and martensite from digital images of carbon steel metallurgical specimens. Digital metallographs were converted to grayscale images and then keypoints from these images were located. An area around each keypoint was extracted as a separate sub-image then reduced to four levels of gray. Using Mathematica , features such as statistical measures of texture (image histograms, entropy, energy, contrast, and homogeneity) and the basic pixel counts from a binarized version of the sub-image were calculated. These features were used as input to a radial basis function neural network. Training and validation data for these features were generated from the sub-images. Using a radial basis neural network (available as a Mathematica add-in) with one output for each phase, it was demonstrated that these features were sufficient to differentiate between all four phases, with a root mean square error of 22.4%. By eliminating cementite from consideration, a root mean square error of 16.2% was achieved.

Biography :

Sara McCaslin completed her Ph.D. in mechanical engineering at The University of Texas at Arlington, and a MS in computer science at The University of Texas at Tyler. She is currently an Assistant Professor of mechanical engineering of UT Tyler, where she pursues research in image processing, finite element analysis, and online education.