Robot vision for automatic inspection of Permanent-magnets
Abstract
This project has been performed in cooperation with Rolls-Royce Marine. They havea new electric motor soon hitting the masses, and there is a hope that the magnetsused in this motor can be automatically quality controlled with computer vision. Thisthesis takes on that challenge, and investigates possible technologies that can be used.Multi-view 3D reconstruction, laser stripe scanning and 2D image comparison techniqueshas been identified as promising candidates. Requirements for the magnets have beeninvestigated, and the solutions that showed potential have been implemented. Opticalcharacter recognition has been identified as an important part of an automated qualitycontrol solution, as one need to register the serial number of each magnet. This hastherefore also been implemented.The different implementations have then been tested and analyzed. It is concluded thatwhile detecting chips from the corners and edges of the magnet could be possible witheither laser scanning or 2D image comparison, the 2D image comparison is simpler andless computationally intensive, and therefore better suited for mass production. The testswhere not able to safely detect the damages to the surface, but this might be possiblewith a better camera. Further work will be needed to conclude on whether this will bepossible. Cracks where not detected with any of the solutions, and is likely the mostdifficult damage to detect. Due to the many difficulties encountered, a proposal has beenmade for a half-automated solution, where a human does visual inspection, the toleranceof the magnet is ensured automatically with a funnel system, and the serial number getsautomatically scanned.For all implementations there are potentials for improvement. Trying different and bettersuited filtering techniques can improve the results of both the laser scanning, the 2D imagetechniques, and the optical character recognition. Using a better camera will also improvethe results, as the camera that has been used has low resolution, high levels of noise, andlacks control of the auto focus, white balance, and aperture. If a 3D solution is chosen,one will also need a solution for how to identify a particular damage in the generated pointcloud, as this has not been investigated in this report.