Automatic visual inspection of metallic surfaces

This thesis is concerned with the objectives of automatic visual inspection of metallic surfaces and involves two major parts. The first part covers three different imaging techniques, gray-level intensity imaging, light sectioning, and photometric stereo. These imaging principles more or less strongly rely on the reflection property of the surface. Therefore, a reflection model for machine vision is introduced. The second part concentrates on the analysis of the gathered data in regard to the detection and classification of surface defects. Additionally, the evaluation of genetic algorithms with a novel encoding scheme and a large number of published sequential feature selection algorithms for selection of the subset of features achieving the best classification rate is included. The genetic algorithms and the adaptive sequential forward floating selection method achieve similar results in performance and computational efficiency. Finally, the results of feature selection and classification of 540 flaw images are presented, whereby different classification approaches such as parametric classifiers, the k-nearest-neighbor decision rule, the naive Bayes classifier, and the tree augmented naive Bayes classifier were compared. For learning the structure of the augmented naive Bayes network a new approach similar to the sequential floating algorithm is presented which achieves a higher classification accuracy than hill climbing search. Basically, the introduced techniques are applied to two fundamentally different applications, whereby the experimental results of both, inspection of high-precision surfaces such as bearing rolls and flaw detection on partially scale-covered steel blocks, are presented. For the inspection of bearing rolls, the surface reflectance properties are modeled and verified with optical experiments. The aim is to determine the optical arrangement for illumination and observation, where the contrast between errors and intact surface is maximized. A new adaptive threshold selection algorithm for segmentation of the irregularities is presented. This algorithm operates on uni-modal as well as bi-modal histograms. For the second application reliable defect detection methods applied on range data are treated. They are based on determination of the distance from the surface profile orthogonal to the surface model. This model is either established by using spline interpolation or alternatively approximated by means of singular value decomposition. That task is complicated by vibrations of the inspected good. Nevertheless, high-speed geometric measurements have been successfully applied to surface inspection of rough steel blocks. (author)