Neural network and skeleton based inspection of surface flaws

This paper explores an inspection method using object´s skeleton pixel counts as neural network inputs to identify surface flaws without constraints on object position and orientation. The associated neural networks have been specifically trained in advance using the skeleton pixel counts generated from the engineering drawings. Because only the pixel counts of an object has been used for inspecting flaws, it tremendously reduces the representative skeleton image data. The significant training time in each epoch is also saved in the application of neural networks. This method can be performed in parallel machines to achieve real time inspections