Performance analysis and optimisation of shape recognition and classification using ANN

Machine vision systems are being increasingly used for sophisticated applications such as classification and process control. Though there is significant potential for the increased deployment of industrial vision systems, a number of important problems have to be addressed to sustain growth in the area of industrial machine vision. Artificial neural networks (ANNs) coupled with machine vision systems offer a new methodology for solving difficult computational problems in many areas of science and engineering. As a consequence, the research work presented in this paper investigates several novel uses of machine vision and ANNs in the processing of single camera multi-positional images for 2D and 3D object recognition and classification. Many industrial applications of machine vision allow objects to be identified and classified by their boundary contour or silhouette. Boundary contour information was chosen as an effective method of representing the industrial component, a composite signature being generated using vectors obtained from the generation of multi-centroidal positions and the boundary pixels. The composite signature can be re-sampled to form a suitable input vector for an ANN. Three different ANN topologies have been implemented: the multi-layer perceptron (MLP), a learning vector quantisation network (LVQ) and hybrid self-organising map (SOM). This method of representing industrial components has been used to compare the ANN architectures when implemented as classifiers based on shape and dimensional tolerance. A number of shortcomings with this methodology have been highlighted, most importantly the identification of a unique sequence start point, vital for rotation invariance. Another problem may arise due to the conflict between the inherent robustness of ANNs when dealing with noise, and classifying components which are similar but display subtle dimensional differences.