| Machine vision has the potential to significantly impact both quality and productivity in computer integrated manufacturing, due to its versatility, flexibility and relative speed. Unfortunately, algorithmic development has not kept pace with advances in vision hardware technology, particularly in the areas of inspection and decision making. The objective of this thesis work is the development of machine vision algorithms for automated inspection of production parts. The shapes of interest are two-dimensional profiles, generated by projecting 3D objects onto a 2D inspection plane, with boundaries that are assumed to be composed of straight-line segments and circular arcs. The inspection system presented in this work consists of three parts in series: segmentation, recognition and analysis. The input of this system is a set of ordered boundary data extracted from the object, and the output includes the identity of this object, and its pose, dimension and profile error.; The first part of this system is to segment the set of ordered boundary data into a desired number (say n) of data subsets, each of which is approximated by a straight-line or an arc. A hybrid procedure, including shape coding schemes and a data fitting technique, is proposed for segmenting the contour into line and arc entities. The object recognition problem is solved using a structural model and neural network computing. The structural model is to represent the objects in terms of the n edge entities, each of which is described by three features: its length, curvature, and relative orientation. Based on this model, the recognition is accomplished by using a multi-layered feed-forward neural network, which classifies the input boundary data (object) into one of the reference shapes. Therefore, the subsequent analysis is a model-based approach, intended to match the input shape, represented by the n data subsets, with a predefined model. Each shape model is defined by four global parameters, based on which n entities of the shape are expressed in implicit form. The matching problem is then solved by finding a best-fit solution between the shape model and the segmented boundary data, using a least-squares technique. The developed algorithms can easily be programmable to inspect different types of shapes by solely changing an input parameter (n), which makes the vision system generic and flexible. |
