Fractographic classification in metallic materials by using computer vision

• We develop an expert system to classify between three failure modes.
• We used Haralick's features, texture energy laws and fractal dimension as descriptors.
• We evaluated two classifiers Artificial Neural Networks and Support Vector Machine.
• The results are comparable with the ones obtained by an expert in this field.
• The study was designed to support failure analysis and as a base for future systems.

The first step in the failure analysis is based on the visual inspection of the fracture surface. This is the base for the development of any fractographic process and it represents the main means for fracture classification. In several occasions, this process is carried out by non-suitable personnel in this area, which increases the chances of generating a wrong classification and, therefore, of negatively altering the results of the entire process. By using artificial vision techniques, this document shows the work done for classifying three types of fractures: brittle sudden, ductile sudden, and progressive due to fatigue, all these in metallic materials in order to promote failure analysis on a fracture surface. The employed descriptors: Haralick's features, energy masks, and the fractal dimension (41, overall), were generated from the Gray Level Co-occurrence Matrix (GLCM), the texture energy laws, and the fractal analysis (respectively) applied to obtained full-scale images, different from other studies using SEM for the acquisition of the data set. The Artificial Neural Networks classifiers and the Support Vector Machine performance were analyzed, finding out that the first one obtained the best results. Such results can be compared to the ones obtained by an expert in this field, with an accuracy percentage higher than 80% for the three types of fractures.