Texture Descriptors for Automatic Classification of Surface Defects of the Hot-Rolled Steel Strip

Abstract

The surface of machined parts is one of the most scrutinised criteria, since it determines their machinability. In this paper, texture descriptors obtained from the Grey Level Co-ocurrence Matrix (GCLM) are used to detect and classify six different types of surface defects of hot-rolled steel strip collected in the NEU dataset. Then, the texture feature vectors are passed to multiple machine learning classifiers to determine the most appropriated one for the dataset, and it is found to be Random Forest. As the features are calculated considering multiple angles, a dimensionality reduction is developed to achieve 94.41% accuracy.