A Hybrid Face Recognition Approach Using GPUMLib



We present a hybrid face recognition approach which relies on a Graphics Processing Unit (GPU) Machine Learning (ML) Library (GPUMLib). The library includes a high-performance implementation of the Non-Negative Matrix Factorization (NMF) and the Multiple Back-Propagation (MBP) algorithms. Both algorithms are combined in order to obtain a reliable face recognition classifier. The proposed approach first applies an histogram equalization to the original face images in order to reduce the influence from the surrounding illumination. The NMF algorithm is then applied to reduce the data dimensionality, while
preserving the information of the most relevant features. The obtained decomposition is further used to rebuild accurate approximations of the original data (by using additive combinations of the parts-based matrix). Finally, the MBP algorithm is used to build a neural classifier with great potential to construct a generalized solution. Our approach is tested in the Yale face database, yielding an accuracy of 93.33% thus demonstrating its potential. Moreover, the speedups obtained with the GPU greatly enhance real-time implementation face recognition systems.


GPU computing, Non-Negative Matrix Factorization, Multiple Back-Propagation, Hybrid systems, Face Recognition


GPU Computing, Machine Learning


Iberoamerican Congress on Pattern Recognition, CIARP 2010, LNCS 6419, Springer, pp. 96-103, November 2010


Cited by

No citations found