Abstract

The automated flaw detection in aluminium castings consists of two steps: a) identification of potential defects using image processing techniques, and b) classification of potential defects into 'defects' and 'regular structures' (false alarms) using pattern recognition techniques. In the second step, since several features can be extracted from the potential defects, a feature selection must be performed. In addition, since the two classes have a skewed distribution, the classifier must be carefully trained. In this paper, we deal with the classifier design, i.e., which features can be selected, and how the two classes can be efficiently separated in a skewed class distribution. We propose the consideration of a self-organizing feature map (SOM) approach for stratified dimensionality reduction for simplified model building. After a feature selection and data compression stage, a neuro-fuzzy method named ANFIS is used for pattern classification. The proposed method was tested on real data acquired from 50 noisy radioscopie images, where 23000 potential defects (with only 60 real detects) were segmented and 405 features were extracted in each potential defect. Using the new method, a good classification performance was achieved using only two features, yielding an area under the ROC curve A x = 0.9976. © Springer-Verlag 2004.