Highly-constrained neural networks with application to visual inspection of machined parts

The authors investigate techniques for embedding domain specific spatial invariances into highly constrained neural networks. This information is used to reduce drastically the number of weights which have to be determined during the learning phase, thus allowing application of artificial neural networks to problems characterized by a relatively small number of available examples. As an application of the proposed technique, the problem of optical inspection of machined parts is studied. More specifically, the performance of a network created according to this strategy which accepts images of the parts under inspection at its input and issues at its output a flag which states whether the part is defective, is characterized. The results obtained so far show that such a classifier provides a potentially relevant approach for the quality control of metallic objects since it offers at the same time accuracy and short software development time.<>