Framework for development of perceptual defect visibility models

Designing an automatic inspection setup for reflective metal object is extremely challenging due to the complex reflection patterns of functional surfaces and pattern variations resulting from the manufacturing process. The process of design requires experience and experimentation with different camera and lighting devices, including their positioning relative to the object, to ensure complete inspection coverage of the object surface. Previous works focus on geometrical visibility or reachability given space requirements, ignoring the perceptual visibility of the defect from the nominal surface patterns. We present a framework for the development of perceptual visibility models, which estimate the probability of defect visibility with regard to the difference of reflection patterns between the defect and the surface. The developed model can be directly used in inspection planning to evaluate the surface coverage of a viewpoint, while considering the perceptual visibility of potentially occurring defects. In our study we collect visibility knowledge using photorealistic rule-based synthetic data. We analyze the defect visibility maps of different surface textures, geometries and geometric defect types. The final model is evaluated using annotated real data, showing the practical alignment of developed models with human estimation.