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.
With escalating pressure being placed on industry to, increase efficiency, improve quality and reduce cost, the need for more flexible and “intelligent” inspection systems has never been greater [1], [2], [3], [4], [5], [6], [7], [8], [9], [10], [11], [12], [13] and [14]. An important milestone in the development of “intelligent” inspection systems has been the rapid growth of computing power in recent years, coupled with the idea that we could successfully emulate the low-level mechanisms of the brain. The human visual system has the ability to recognise an object despite changes in the object's position in the input field, its size, or its orientation. For many industrial applications involving classification of components machine vision systems must also have this ability. A simple approach to recognition/classification of industrial components is to segment the image into two major stages, these being extraction and the actual analysis and processing stages. Many current and recent artificial neural network (ANN) classifiers require a 2D shape to be presented in a fixed position, orientation and dimension. Methods have been proposed for rotation normalisation where the network is trained with a set of rotated signatures. The network was then capable of classifying the shapes/signatures in all possible angles. However, rotation of shapes may produce disconnected decision regions leading to problems in network convergence.
The programs which attempt to emulate the low-level mechanisms of the brain are neural networks and fuzzy logic. Over the years there has been increased interest in the synergistic combination of machine vision, neural networks and fuzzy systems in the classification of industrial workpieces [15], [16], [17], [18], [19], [20], [21], [22], [23], [24], [25], [26], [27], [28] and [29]. The synergism of machine vision, neural networks and fuzzy logic seems natural. This paper describes part of a continuing program of work developing an artificial vision system, applicable to real-time recognition/classification of industrial components.
The ability of the single start point (SSP) method and the ‘dynamic’ positional scheme to successfully identify objects that display a range of shape signatures has been demonstrated, but the complexity of the problems where they are capable of offering a perfect solution has not been totally quantified. The number of classes the network has to cope with affects the complexity of the problem and the training time increases dramatically. Whilst the Euclidean distances for a particular shape are equally spaced around the boundary, it is possible for certain objects that it may not necessarily represent the features in the most effective way. The shape may possess relatively large smooth regions on the boundary where the use of more than one or two samples could be wasteful, on the other hand the object may have concentrated regions of greater detail where the number of samples needs to be greater in order to represent the important features. However, the techniques employed within this research produced acceptable results for the given application. One of the areas where further investigation would be of interest is that of identifying 3D objects from a series of 2D images.
