Impact of Object Extraction Methods on Classification Performance in Surface Inspection Systems

In surface inspection applications the main goal is to detect all areas which might contain defects or unacceptable imperfections, and to classify either every single 'suspicious' region or the investigated part as a whole. After an image is acquired by the machine vision hardware, all pixels that deviate from a pre-defined 'ideal' master image are set to a non-zero value, depending on the magnitude of deviation. This procedure leads to so-called ``contrast images'', in which accumulations of bright pixels may appear, representing potentially defective areas. In this paper, various methods are presented for grouping these bright pixels together into meaningful objects, ranging from classical image processing techniques to machine-learning based clustering approaches. One important issue here is to find reasonable groupings even for non-connected and widespread objects. Generally, these objects correspond either to real faults or to pseudo-errors that do not affect the surface quality at all. The impact of different extraction methods on the accuracy of image classifiers will be studied. The classifiers are trained with feature vectors calculated for the extracted objects found in images labelled by the user and showing surfaces of production items. In our investigation artificially-created contrast images will be considered as well as real ones recorded on-line at a CD imprint production and at an egg inspection system.