Empirical comparison of defect classifiers on specular surfaces

Today the inspection of specular surfaces, especially in the automobile sector, is often done by humans. An automated inspection would be preferable for reasons such as reproducibility, reliability, and objectivity. However it is problematic to replace humans by machines in this field. The main reasons for this are their greater flexibility for changes in the production process and their ability not only to find defects but to decide whether a customer would complain about those defects. With the deflectometric principle, there is a measurement method for specular surfaces that is fast and accurate enough to compete with humans. Open problems are the necessary expenses for the parameterization of the defect detection and the missing link to the human perception of defects. The first problem is addressed in this paper. An overview of methods capable of detecting and classifying defects of different shapes and scales on unknown surface shapes is given. Then all methods are compared empirically on real measurement data