Imbalanced Learning Ensembles for Defect Detection in X-Ray Images.

This paper describes the process of detection of defects in metallic pieces through the analysis of X-ray images. The images used in this work are highly variable (several different pieces, different views, variability introduced by the inspection process such as positioning the piece). Because of this variability, the sliding window technique has been used, an approach based on data mining. Experiments have been carried out with various window sizes, several feature selection algorithms and different classification algorithms, with a special focus on learning unbalanced data sets. The results show that Bagging achieved significantly better results than decision trees by themselves or combined with SMOTE or Undersampling.