Face Verification using Gabor Filtering and Adapted Gaussian Mixture Models Laurent El Shafey1,2, Roy Wallace1 and S´ebastien Marcel1 1

Idiap Research Institute, Switzerland 2´ Ecole Polytechnique F´ed´erale de Lausanne, Switzerland 4. System Overview

1. Abstract I

I

I

I

Face authentication is difficult because of the high-variability of face images This work proposes an approach combining the strengths of both Gabor-based features and GMM modelling In particular, we model each output of a Gabor filterbank separately, training a set of highly specialized classifiers The proposed system demonstrates up to 52% relative improvement in verification error rate compared to a standard GMM approach

Figure: Overview of the proposed Gabor DCT-GMMsystem

We consider the outputs of the Gabor filtering separately, training a highly specialized model for each particular frequency subband. I Score fusion is performed using the sum rule. I

We compare our novel Gabor DCT-GMM system to the baselines referred to as Eigenfaces, DCT-GMM and LGBPHS [1]. I For all systems, images are preprocessed using Tan and Triggs method [2]. I

2. Gabor Filtering I

Gabor filters have invariance properties to translation, scale and rotation. We expect that this will provide robustness to pose, illumination and expression. Filtering is performed in the frequency domain for efficiency. I In the Fourier domain, Gabor filters are Gaussian. 2 − πf2 [(u 0−f)2γ2+v 02η2] Ψ2(u, v) = e , u 0 = u cos (θ) + v sin (θ) , (1) 0 v = −u sin (θ) + v cos (θ) . I We minimize redundant information between filters by setting. r   (k + 1) ln p1 , γ= (k − 1) π 1 r η= . (2)
π π2 1 tan 2N − γ2 ln ( p1 )

5. Experiments

I

p : Maximum overlap level N : Number of filters I

I

We only use the magnitude of the filters, and can restrict ourself to half of the frequency-space.

5.1 Databases I

We use challenging, publicly-available databases and protocols with separate training, development and test sets to allow for unbiased evaluation.

(a) BANCA database: controlled, degraded, and adverse scenarios.

Figure: Representation of a grid of Gabor filters in the frequency domain.

(b) MOBIO database: still face protocol.

We noticed that the design of the filterbank has a huge impact on the performance.

5.2 Results on BANCA and MOBIO

Frequency (f) Gabor DCT-GMM LGBPHS 0.05 16.67 15.64 0.15 3.66 8.54 0.25 4.95 5.90 0.35 11.28 7.87 0.45 14.80 16.67 Table: Verification performance (HTER) of Gabor-based systems by frequency, fusing over the 8 orientations on the tuning set (group g1 of BANCA G)

3. Modelling the Face with Gaussian Mixture Models

(c) BANCA

A feature vector from each block

(d) MOBIO

Figure: Verification performance (HTER) of the baseline systems (Eigenfaces, DCT-GMM, LGBPHS) and of the novel Gabor DCT-GMM system on the different protocols of the BANCA and MOBIO databases

6. Conclusions and Future Work The proposed system demonstrates substantial improvements compared to the standard DCT-GMM approach (52% relative HTER improvement on the G protocol of BANCA). I The proposed system outperforms a state-of-the-art LGBPHS technique on four of the seven BANCA face verification protocols, and on the male protocol of the MOBIO database. I Future work will investigate ways to be more robust in case of unmatched training and testing conditions, as well as the use of a more complex Bayesian approach to combine information from the filters. I

Input image

Image blocks

2D-DCT features extracted from each part of the face independently ⇒ naturally robust to occlusion, local transformation and face mis-localisation. I Modelling: The distribution of feature vectors for each client is modelled by a Gaussian mixture model (GMM). I Scoring: Calculate a log likelihood ratio for an image Ot between the model of the claimed client identity, si, and a universal background model (UBM), m: K X
k k h (Ot, si) = log(p(ot | si)) − log(p(ot | m)) . I

k=1

References [1] W. Zhang, S. Shan, W. Gao, X. Chen, and H. Zhang, “Local Gabor Binary Pattern Histogram Sequence (LGBPHS): A novel non-statistical model for face representation and recognition,” in IEEE International Conference on Computer Vision (ICCV), vol. 1, 2005, pp. 786–791. [2] X. Tan and B. Triggs, “Enhanced local texture feature sets for face recognition under difficult lighting conditions,” IEEE Transactions on Image Processing, vol. 19, no. 6, pp. 1635–1650, 2010.

http://www.idiap.ch, http://www.epfl.ch, http://bbfor2.net

{Laurent.El-Shafey, Roy.Wallace, Sebastien.Marcel}@idiap.ch