http://digital-library.theiet.org/content/conferences/10.1049/ic.2013.0254

1. An Evaluation of Denoising Algorithms for 3D
Face Recognition
Mehryar Emambakhsh, Jiangning Gao and Adrian Evans
Department of Electronic and Electrical Engineering
University of Bath

2. Outline
• Introduction
• Noise in 3D data
• Denoising methods
• Research questions
• Recognition pipeline
• Classification algorithms
• Experimental results
• Conclusion

3. Noise in 3D data
• Randomly added to the data and inevitable!
• On the face surface it produces
• Missing data
• Spike noise
• Holes

4. Noise in 3D data
• Noise can significantly affect:
• Pose correction
• Landmarking algorithms and face cropping
• In addition, noise can:
• Insert outliers in the feature space  Normalisation
• Degrade the classification/matching performance  Probe vs.
gallery variation  The within/between-class scatter

5. Denoising methods
• The main purpose is to reduce the noise level, while
preserving the information required for the recognition
• Several methods have been used:
• Mean filtering
• Median filtering
• Gaussian filtering
• Weiner filters
• Wavelets
• Non-linear diffusion

6. Denoising methods
• Research questions:
• Which denoising method performs better for baseline holistic
algorithms?
• How to choose the best parameters for
• Mask size?
• Variance?
• Wavelet type and levels?
• Diffusion equation parameters?
• How sensitive different denoising methods are compared to
each other?
• Does aggressive denoising reduce the recognition
performance?

7. Recognition pipeline
• Resampled to 0.5 mm/pixel
• Spherical intersection for cropping
• PCA-based alignment
• Normalised to [0, 1]
A.S. Mian, M. Bennamoun, and R. Owens. An efficient multimodal 2D-3D
hybrid approach to automatic face recognition. IEEE Trans. Pattern Anal.
Mach. Intell., 29(11):1927–1943, 2007.

8. Recognition pipeline
• Classification algorithms
• Multi-class SVM
• Principal Component Analysis (PCA) or Eigenfaces
• Kernel Fisher’s Analysis (KFA)
• Linear Discriminant Analysis (LDA) or Fisherfaces

9. Recognition pipeline
• Classification algorithms (cont.)
• Probabilistic Neural Network (PNN)
• K- Nearest Neighbour (KNN)
• Bootstrap Aggregation Decision Trees (TreeBagger)

10. Experimental results
• Dataset
• Face Recognition Grand Challenge (FRGC v2.0). 3 Folders:
• Spring 2003  noisier, both spike and white noise additive
noise
• Fall 2003
• Spring 2004
• For Spring 2003 folder:
• Minolta Vivid 900/910 laser for capturing
• The classes with at least four samples are used  119 subjects
and 661 samples
• Two samples for training and the remaining for test

11. Experimental results: Mean filtering
C1: Multi-SVM
C2: PCA
C3: KFA
C4: PNN
C5: KNN
C6: TreeBagger
C7: LDA

12. Experimental results: Gaussian filtering
C1: Multi-SVM
C2: PCA
C3: KFA
C4: PNN
C5: KNN
C6: TreeBagger
C7: LDA

13. Experimental results: Median filtering
C1: Multi-SVM
C2: PCA
C3: KFA
C4: PNN
C5: KNN
C6: TreeBagger
C7: LDA

14. Experimental results: Weiner filtering
C1: Multi-SVM
C2: PCA
C3: KFA
C4: PNN
C5: KNN
C6: TreeBagger
C7: LDA

15. Experimental results: Non-linear diffusion

16. Experimental results: Non-linear diffusion

17. Experimental results: Non-linear diffusion
C1: Multi-SVM
C2: PCA
C3: KFA
C4: PNN
C5: KNN
C6: TreeBagger
C7: LDA

18. Experimental results: Wavelets
• KFA classification on denoised faces using wavelets
• Recognition performance found for dwt denoising with 1
to 10 levels

19. Experimental results: Wavelets
• There is again a slight upward trend in the average
curve

20. Experimental results: KFA classification
MEY (Discrete Meyer), AF (Average or mean filtering), GF (Gaussian
filtering), DIFF (non-linear diffusion),WEI (Weiner filtering, MED
(median filtering), Mian et al. and UFF (unfiltered faces)

21. Conclusion
• The performance of different denoising algorithms on
various holistic face recognition methods was
evaluated.
• Experimental results show:
• Aggressive denoising does not necessarily reduce the
recognition performance
 Low frequency components are more informative for face
recognition
• The optimum mask sizes are significantly larger than those
previously used
• Median filtering (16.5×16.5 mm2) produced the highest
recognition rate (98.35%) for the KFA classifier

22. Conclusion
• The proposed method is dataset independent
• Photometric stereo
• 4D datasets
• Can be used to tune the parameters of any biometric
system’s denoising algorithm

23. Thank you!