Template Guided Visual Inspection A. Noble, V.D. Nguyen, C. Marinos, A.T. Tran, J. Farley, K. Hedengren, J.L. Mundy GE Corporate Research and Development Center P.O. Box 8, 1 River Road Schenectady, NY. 12301. USA. Abstract. In this paper we describe progress toward the development of an X-ray image analysis system for industrial inspection. Here the goal is to check part dimensions and identify geometric flaws against known toler- ance specifications. From an image analysis standpoint this poses challenges to devise robust methods to extract low level features; develop deformable parameterized templates; and perform statistical tolerancing tests for ge- ometry verification. We illustrate aspects of our current system and how knowledge of expected object geometry is used to guide the interpretation of geometry from images. 1 Introduction Automatic Visual inspection is a major application of machine vision technology. How- ever, it is very difficult to generalize vision system designs across different inspection applications because of the special approaches to illumination, part presentation, and image analysis required to achieve robust performance. As a consequence it is necessary to develop such systems Mmost from the beginning for each applicatign. The resulting development cost prohibits the application of machine vision to inspection tasks which provide a high econonfic payback in labor savings, material efficiency or to the detection of critical flaws involving human safety. The use of Computer Aided Design (CAD) models has been proposed to derive the necessary information to automatically program visual inspection [16, 2]. The advantage of this approach is that the geometry of the object to be inspected and the tolerances of the geometry can be specified by the CAD model. The model can be used to derive optimum lighting and viewing configurations as well as provide context for the application of image analysis processes. On the other hand, the CAD approach has not yet been broadly successful because images result from complex physical phenomena, such as specular reflection a~ld mu- tual illumination. A more significant problem limiting the use of CAD models is that the actual manufactured parts may differ significantly from the idealized model. During product development a part design can change rapidly to acconmmdate the realities of manufacturing processes and the original CAD representation can quickly become obso- lete. Finally, for curved objects, the derivation of tolerance offset surfaces is quite complex and requires the solution of high degree polynomial equations [3]. An alternative to CAD models is to use an actual copy of the part itself as a reference. The immediate objection is that the specific part may not represent the ideal dimensions or other properties and without any structure it is impossible to know what attributes of the part are significant. Although the part reference approach has proven highly successful in the case of VLSI photolithographic mask inspection [5, 14] it is difficult to see how to extend this simple approach to the inspection of nmre complex, three dimensional, manufactured parts without introducing some structure defining various regions and boundaries of the part geometry. The major problem is the interpretation of differences