Automatic crack detection based on machine learning and DVC
Résumé
A voxel-based crack detection method is proposed from DVC results and a simple machine learning algorithm. The method, that requires almost no input and manual fine tuning, exploits observable characteristic of brittle cracks and independant of the material and the experimental procedure. Two applications are performed on two different materials whose cracks were imaged on a tomograph and highlight the efficiency and simplicity of the proposed method. Introduction Images play an important role in the mechanical characterization of materials (e.g., DIC). In particular, the matter of identifying the presence of a crack and its subsequent extraction has been widely explored and is usually based on a ad hoc sequence of classical segmentation methods (e.g., simple threshold, Otsu's method, etc) followed by morphological operations (i.e., erosion and dilation) so as to "clean" the segmentation. Two main inconveniences arise with such procedure. First, the custom nature of such procedures means that they need to be performed for each use case. And second, the inherent subjectivity present in the choice of the required parameters (e.g., thresholds, size of erosion kernel, etc.). The challenges imposed by this latter were addressed in references [1, 2] with the use of observables based on features of the material (and its behavior). This approach allows removing the subjectivity in the choice of a threshold value, since it is chosen so as to better approach the known observable [1]; as well as allowing to judge the uncertainty that its related to the procedure, by observing the variability of various micro-mechanical simulations guided by the segmentation procedure [2]. The goal of this research is to provide a methodology that simultaneously addresses both issues previously highlighted.
Origine : Fichiers produits par l'(les) auteur(s)