Identification and characterization of fracture in metals using machine learning based texture recognition algorithms

Abstract

Manual identification of brittle and ductile fracture regions in fractographic images of metals is cumbersome, time-consuming, and can be a subjective process. A supervised machine learning classifier in conjunction with a texture recognition algorithm is employed in this study to automatically identify the fracture type and evaluate their area fractions in fractographic images. The texture is a unique visual characteristic possessed by an object in an image that distinguishes it from the other objects. In the context of this study, both brittle and ductile fracture is assumed to possess unique textures. To quantify texture, Local Binary Pattern (LBP) texture quantification algorithm is employed in this study and the corresponding set of statistical metrics referred to as textural features are evaluated from both ductile and brittle fracture training images. These textural features are then used to train Linear Discriminant Analysis machine learning classifier whose performance is assessed using cross-validation technique. The trained algorithm is then deployed to identify the brittle and ductile fracture regions in fractographic images that are not used for training purposes. In this study, five different fractographic images of structural steels are used for validation purposes, and the fracture type is successfully identified.