Visual inspection of axisymmetric parts

Flexible manufacturing systems is a field of growing interest for applications of automated visual inspection systems. This paper addresses the problem of automated blemish detection in axisymmetric engineering parts. Typically these parts are visually analysed by human operators to detect anomalies. This is a tedious and time-consuming process that is subject to variables as image quality, and inspector expertise and fatigue. In order to reduce inspection time and improve detection reliability, a system that consists of a sequence of image processing steps and a classification step is proposed. The image processing techniques required for this problem are considered. Four different classification schemes, including Fisher's method and a neural network are used and compared. This choice of structure is based on earlier work which proposes that although many image processing steps may be implemented with neural nets, it is only the classifier that is currently optimally implemented in this way. Fisher's linear discriminant may provide a useful tool in the selection of the neural network architecture. The proposed system is able to inspect 'real-world' axisymmetric parts in simulated real-time and is robust to changes in translation, orientation and scale.