Detection of flaws in aluminium castings: a comparative study between generative and discriminant approaches

Automatic anomaly detection has become a key issue in many engineering applications due to the increasing amount of data in need of analysis. Addressing this kind of task using pattern recognition methods requires a proper design of the learning strategy, given the reduced amount of flawed cases available for training compared to that of normal instances, which has been shown to hinder the performance of traditional classification algorithms. Moreover, positive examples are often costly and hard to collect, which may prevent the use of traditional discriminant approaches such as artificial neural networks. In this paper, we compare two well-known generative and discriminant pattern recognition algorithms in the task of flaw detection in aluminium castings and show that defects can be accurately identified without prior knowledge of positive cases, using only information of regular structures, achieving a geometric mean of over 0.9.