wPFP-PCA:
Weighted Parallel Fixed Point PCA Face Recognition
Chakchai So-In
and Kanokmon Rujirakul
Department of
Computer Science, Khon Kaen University, Thailand
Abstract: Principal Component Analysis (PCA) is one
of the feature extraction techniques, commonly used in human facial recognition
systems. PCA yields high accuracy rates when requiring lower dimensional
vectors; however, the computation during covariance matrix and eigenvalue
decomposition stages leads to a high degree of complexity that corresponds to the
increase of datasets. Thus, this research proposes an enhancement to PCA that
lowers the complexity by utilizing a Fixed Point (FP) algorithm during the
eigenvalue decomposition stage. To mitigate the effect of image projection
variability, an adaptive weight was also employed added to FP-PCA called wFP-PCA.
To further improve the system, the advances in technology of multi-core
architectures allows for a degree of parallelism to be investigated in order to
utilize the benefits of matrix computation parallelization on both feature
extraction and classification with weighted Euclidian Distance optimization.
These stages include parallel pre-processor and their combinations, called
weighed Parallel Fixed Point PCA wPFP-PCA. When compared to a traditional PCA
and its derivatives which includes our first enhancement wFP-PCA, the
performance of wPFP-PCA is very positive, especially in higher degree of recognition
precisions, i.e., 100% accuracy over the other systems as well as the increase
of computational speed-ups.
Keywords: Face recognition, FP, parallel face recognition, parallel euclidian, parallel PCA, PCA.
Received December 27, 2014; accepted May 21, 2014