الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 122 |  |  |
| | | from | 122 | | |
Abstract
Surface reconstruction is a necessary step in Reverse Engineering (RE). It recovers the 3D shape of the part from point cloud acquired from its surface. The recovered shape can be then used by CAD/CAM applications for modifying and manufacturing the part. Surface reconstruction can be done by: polygonal mesh, or Constructive Solid Geometry (CSG), or parametric surface fitting. Parametric surfaces provide the most sophisticated and most accurate representation of freeform surfaces. They include: Bezier, B-spline, and NURBS (Non Uniform Rational B-spline) surfaces. NURBS surfaces are widely used due to their flexibility and local control ability. Also they can be used to model both of freeform and analytical shapes. In this thesis, a new type of neural networks called Rational B-spline Neural Networks (RBNN) is designed to approximate point clouds to NURBS curves and surfaces. NURBS approximation methods include: Least Squares Methods, Genetic algorithms, simulated annealing, and Particle Swarm Optimization. Compared to these methods, RBNN do not require parameters or knots optimization. Training of the network is achieved using variable training rate backpropagation algorithm. To validate the networks, MATLAB programs are made and actual parts are digitized on CMM (Coordinate Measuring Machine). By entering the obtained points to the developed program, it is found that better approximation error can be obtained provided that the number of control points is properly selected. Also, the networks can be applied on small and large volume point cloud. Time consumption in training is proportional to the volume of point cloud. For 400~600 points, RBNN require about 20 seconds until reaching the desired error. For 2150 points, RBNN require about 60 seconds until reaching the desired error.