Automatic Machine Vision Calibration Using Statistical and Neural Network Methods Lyndon N. Smith, Melvyn L. Smith Faculty of Computing, Engineering and Mathematical Sciences (CEMS) University of the West of England Bristol, BS16 1QY, UK Abstract A methodology for the correction of errors in machine vision images is presented. The regression and neural network techniques studied are considered to be complimentary rather than competitive. Neural networks have been identified as particularly useful for precise modelling of non-linear responses, and offer the additional benefits of being non-prescriptive and generally applicable to factors such as radial lens distortion and minor camera miss- alignments. The combination of these modelling techniques within automated program control is suggested as an approach for straightforward and accessible machine vision calibration. 1. Introduction Machine vision is a technology that can be employed for implementation of in-process measurement and inspection for a multitude of industries [29]. Information relating to critical process parameters, such as surface features [30-34], or specified dimensions of manufactured components, can be combined with techniques such as Statistical Process Control [12] and Adaptive Control [39] to assist with reductions in manufacturing variations. Such reductions result in real cost savings (even when the variation is within the specification for the product), the magnitude of which can be calculated through use of the Taguchi Loss Function [11]. Machine vision can also offer a number of additional benefits over traditional measurement techniques such as probing and manual measurement. For example, the high speed at which machine vision inspection can be implemented means that often all of the products made can be inspected without interrupting production. The productivity benefits obtainable through machine vision inspection are therefore strong motivators for introducing the technology to the production environment. However, bearing in mind the potential of this technology, the current number of installations is perhaps less than would be expected [23]. There may be a number of reasons for this, including general unfamiliarity with the techniques, or perceptions that the apparatus and knowledge needed are complex and therefore expensive. In reality the equipment needed today for implementation of machine vision is relatively inexpensive and easy to obtain. It is however often true that precise installation and calibration can be a relatively lengthy and therefore expensive process. Consequently, research into methods for simplification of this aspect of machine vision will enable more widespread application of the technology with the attendant economic benefits for the system suppliers as well as the customers. One approach for achieving this is through an automatic calibration that employs statistical and neural network methods.