Uma análise de funções Wavelet para a tarefa de reconhecimento facial

Abstract

Face recognition is an important research topic because of its wide range of applications, such as surveillance, biometrics and control access. Waveletfaces is a feature extraction technique that can improve the accuracy rate of a face recognition system. Nonetheless, its results may deeply vary depending on the wavelet function, decomposition level, classifier, other dimensionality reduction techniques used along with Waveletfaces, as well as the face database. To determine whether there is a subset of such items that can improve Waveletfaces, we performed an extensive comparison. We evaluate 4 dimensionality reduction methods using Waveletfaces, 106 different wavelet functions, 5 decomposition levels, 4 classifiers, and 5 face data sets. The combination of all these elements results in 42,400 scenarios at most. We used the confidence interval hypothesis test to compare the accuracy rate of each scenario with the maximum one within each data set. We determine that a few wavelet functions, such as some from the Reverse Biorthogonal family, can greatly improve classification accuracy. It is also shown that the Nearest Neighbor classifier performs well on all five databases. Moreover, other elements are very related to database issues.