Neural network visual inspection with boundary learning based on the distance index in input space

In the field of electronics device assembly, miniaturization of components, denser packing of boards, surface mounting technology, and highly automated assembly equipment make the task of inspecting the defects of soldering joints in the electronics products more critical and more difficult for humans. The automated inspection systems are required for the stable inspection of products. One of the approaches achieving flexible information processing is the neural network visual inspection system. In our study, the number of input order and boundary learning method are investigated for speedy and correct judgments. Firstly, the size of input images are optimized and principal component analysis is confirmed to be effective for fast inspection with keeping judgment rates. Secondly, we proposed a boundary learning method based on the distance index to select samples in the boundary area in the learning space where the samples are difficult to be judged. And we also achieved cutting down the undeterminable rate and improving correct judgment rate by combination of multiple results obtained with boundary learning methods.