Hole Filling Algorithm in Surface Reconstruction Based on Radial Basis Function Neural Network

Abstract:

Aiming at hole filling in points cloud data reconstruction, a novel neural network arithmetic was employed in abridged points cloud data surface reconstruction. Radial basis function neural network and simulated annealing arithmetic was combined. Global optimization feature of simulated annealing was employed to adjust the network weights, the arithmetic can keep the network from getting into local minimum. MATLAB program was compiled, experiments on abridged points cloud data have been done employing this arithmetic, the result shows that this arithmetic can efficiently approach the surface with 10-4 mm error precision, and also the learning speed is quick and hole filling algorithm is successful and the reconstruction surface is smooth. Different methods have been employed to do surface reconstruction in comparison, the results illustrate the error employed algorithmic proposed in the paper is little and converge speed is quick.

• Periodical: Key Engineering Materials (Volumes 392-394)
• Pages: 750-754
• DOI: https://doi.org/10.4028/www.scientific.net/KEM.392-394.750
• Online since: October 2008
• Authors: X.M. Wu, Gui Xian Li, W.M. Zhao
• Keywords: Abridged Points Cloud Data, Hole Filling Algorithm, Radial Basis Function Neural Network (RBFNN), Simulated Annealing Arithmetic, Surface Reconstruction

References

[1] G. Casciola, D. Lazzaro, L.B. Montefusco and S. Morigi: Numerical Algorithms, Vol. 39 (2005), pp.289-305.

[2] Y. Xing, Y. Wang and X.Y. Ruan: High Technology Letters, Vol. 7 (2001) No. 3, pp.49-54 (In Chinese).

[3] Y.G. Fan, X.M. Liu and J. Chen: Computer Engineering and Application, (2005) 11, pp.66-69 (In Chinese).

[4] K. Wang, C.M. Zhang: CAD&CG, Vol. 10 (1998) No. 3, pp.193-199 (In Chinese).

[5] D. Shi, D.S. Yeung and J. Gao: Neural Networks, Vol. 18 (2005), pp.951-957.

[6] P. Gu, X. Yan: Computer-Aided Design, Vol. 27 (1995) No. 1, pp.59-64.

5 10 15 20 -3 -2 -1.

[1] [2] [3] x 10 -3 epochs （error mm) Error surface of RBFNN.

10 20 30 40 -6 -4 -2.

[2] [4] [6] x 10-4 epochs error(mm) Error surface of method in paper (a) Method in paper （b）RBFNN Fig. 7 Network capability comparison.