A Novel Machine Vision Algorithm for a Fast Response Quality Control System

We are investigating the design of machine vision algorithms for real time inspection of filter components for quality assurance. Filter components are rigid 3D objects with predefined geometry so that a good deal of knowledge can be incorporated into the system design. However, while the objective is to reason about 3D structure, current machine vision techniques only allow the acquisition of 2D object descriptions and this is normally referred to as a 3D-2D correspondence problem. We have investigated geometrical techniques in 2D and in 3D and have developed a novel method to analyse rigid body transformations that have been successfully applied to machine vision calibration problems. In this paper, we first describe a geometrical analysis of image correspondence applied to 2D rigid body transformations. We then develop a novel calibration algorithm that, given a 3D model of an object and a set of 3D-2D image correspondence points allows the calibration of all transformation parameters of interest. The calibrated parameters can then be used to verify the physical dimensions of the object under inspection. For a comparative analysis, we also develop a calibration algorithm based on epipolar geometry applied to the same task. Experimental results have shown that our novel algorithm performs much better than the algorithm based on epipolar geometry and that it is well suited to the real time requirements of the task.