An Experiment on Defect Detection in Active Thermography using Classifier Fusion

This paper presents an experimental study on classifier fusion applied to the detection of defects in signals from infrared thermography non-destructive evaluation. The fusion stage combines the results of individual classifiers in order to improve the overall result of classification in terms of accuracy and variance. Several features were extracted from a set of the thermography signals measured in a composite material. Those features were the input to the following classifiers: linear discriminant analysis (LDA); random forest (RDF); support vector machine (SVM); Bayesian classifier (ASPEC); quadratic discriminant analysis (QDA); naive Bayes (NB); logistic regression (LogReg); and independent component analysis mixture model (ICAMM). Two fusion methods were implemented: the Mean, and alpha integration (a-INT). The results show that a-INT outperforms the other classification methods in terms of precision and balanced error (BER) indicators as well as receiver operating characteristic (ROC) curves.