Abstract:
Three dimensional (3D) face recognition has the potential to achieve better accuracy than its two
dimensional (2D) counterpart by measuring geometry of rigid features on the face. This avoids
such pitfalls of 2D face recognition algorithms such as change in lighting, different facial
expressions, make-up and head orientation. 3D face recognition accuracy can also deteriorate due
to missing data, large pose variation, occlusion, facial expression deformations and time delay
resulting based on acquisition technique. For any 2D or 3D face recognition system, feature
extraction and classification play a critical important role because face biometric depends on these
two stages to make significant progress. In order to develop a robust feature extraction algorithms
that maintains a high discriminative capability, a block-wise statistics based feature extraction
scheme was adapted in this research. Specifically, a 3D face region of interest was divided into
uniform blocks in the first stage and a histogram of surface types was extracted from each block
in the second stage. Histograms from all the blocks were then concatenated to form the final feature
vector. In the classification stage, a novel Kernel Superimposed Sparse Parameter (KSSP)
classifier was developed for the 3D face recognition. The KSSP utilizes the high-dimensional
nonlinear information instead of the linear information in the SSP classifier which was helpful for
classification. The KSSP classifier first used the test sample vector and whole train space to
calculate the global sparse regression parameters. Then, the KSSP score was computed as the
superposition of the sum of the sparse parameters belonging to the same class iteratively. The final
largest KSSP score of the class was used to classify the test sample. Exponential chi-square kernel
was used to exploit a SSP to its nonlinear counterpart. The implementation of the proposed method
was carried out in Matlab environment on window 10 operating system. Other classification
methods were implemented using the block-wise Statistics based Features extraction scheme. The
experiment revealed that the approach achieved a identification rate of 100% at a reasonable time cost of 129.01s and outperformed Label Consistent K-Singular Value Decomposition (LC-
KSVD), Gabor feature based Robust Representation Classification (GRRC) and Kernel Gabor
feature based Robust Representation Classification (KGRRC) with 0.71%, 7.67%, and 4.89%
respectively. The high recognition rate makes it suitable for large-scale identification application.
