Towards Surface Inference in Industrial Inspection

Automated product inspection play an important role in today's manufacturing process, and therefore, the design of optimized and precise measurement setups are a requirement for efficient product quality assurance. Due to the high dimensionality of the design space, a manual choice of the geometrical and optical parameters is associated with high costs, tedious experimental work, and often non-optimal results. Thus, automatic planning methods which seek to optimize the setup degrees of freedom for a particular measurement are of special importance in this field. For automatic evaluation of an inspection, there exist typical evaluation metrics including but not limited to, the measurement uncertainty and the scan resolution. However, it is often not trivial how to combine different optimization criteria to optimize the setup based on the requirements. For example, it is not obvious how to compare an the result of an inspection with a high lateral resolution and a high uncertainty against another inspection, with a low lateral resolution but precise measurements. We propose to fuse the metrics through a probabilistic surface inference to quantify the amount of information gained by a specific setup configuration. To this end, we model the surface by a Gaussian random process and introduce a local surface inference method based on the local surface orientation. The measurement points delivered by a laser triangulation setup are simulated using real-time graphical simulations, and the uncertainty of single point measurements are estimated. This data is further used as input to the local inference method. The inference results are demonstrated for the inspection of the intake manifold of a cylinder head.