An approach to fast detection of defects in metal surfaces

In this thesis an approach for the fast detection and classification of defects in digitized images of surfaces with known intensity distributions is presented. The reflectance response of specular planar surfaces is modeled. It is shown that by proper selection of lighting, non-linear intensity distributions of metal surfaces can be approximated by linear intensity distributions. Two kinds of defects, streaks and speckles, are considered for detection and classification. The approach presented in this thesis has two parts: defect extraction and defect classification. The defect extraction unit generates defect signatures that are input to the classifier. A multilayer perceptron applying the back-propagation rule is utilized for classification of the defect signatures. When applying the proposed approach, more than 95% of the defects are detected and 90% of the detected defects are classified correctly.