Object Recognition in Deviation Images for Fault Detection - A Comparison of Methods

In this paper several algorithms are exploited for the purpose of object recognition in so-called deviation images. These images are obtained when calculating the deviation of a newly recorded image with its master and serve as a basis for a successful discrimination between good and bad images (representing fault-free and faulty system states). The applied object recognition approaches are connected components with and without morphology, iterative prototype-based clustering and hierarchical clustering with some extensions. They will be applied in a red line throughout the whole paper based on different structures appearing various types of deviation images. Furthermore, in the evaluation section it will be outlined, how the different object recognition algorithms perform when applied in an image classification framework for extracting objects out of 20000 images. The performance is measured by observing the impact of the object recognition approaches on the miss-classification rate of a decision-tree based classifier, which is built up based on some features extracted out of the recognized objects.