Combinação de múltiplos classificadores para reconhecimento de face humana

Abstract

Lately, the human face object has been exploited by the advent of systems involving biometrics, especially for applications in security. One of the most challenging applications is the problem of human face recognition, which consists of determining the correspondence between an input face and an individual from a database of known persons. The process of face recognition consists of two steps: feature extraction and classification. In the literature of face recognition, different techniques have been used, and they can be divided into holistic techniques (implicit feature extraction), feature-based techniques (explicit feature extraction) and hybrid techniques (involving the two previous). In many articles, holistic techniques have proved to be most efficient and generally they involve methods of statistical pattern recognition as Principal Component Analysis (PCA), Independent Component Analysis (ICA), Linear Discriminant Analysis (LDA), Support Vector Machines (SVM), and Neural Networks. For problems such as human face recognition in digital images, a crucial point is the ability to generalize. The solution for this problem is complex due to the high dimensionality of data and the small number of samples per person. Using a single classifier would reduce the ability of recognition, since it is difficult to design a single classifier in these conditions that capture all variability that span the human faces spaces. Thus, this work proposes to investigate the combination of multiple classifiers applied to the problem of face recognition, defining a new scheme to resolve this problem, varying the feature extraction with PCA and some its variants and LDA, K-Nearest Neighbor (K-NN) and Maximum Likelihood (MaxLike) classifiers and several trainable or not trainable methods for combining classifiers. Still, to mitigate the problem of small sample size (SSS), a technique for regularizing the covariance matrix was used. Finally, to assess the classification performance, Holdout and Resubstitution methods were used to partition the data set and the Kappa coefficient and Z and T statistics were used to measure the performance of the proposed scheme. From the experiments it was concluded that the best sub-schemes were the RBPCA/MaxLike-PCA/NN-KL5/NN classifiers combined by the Majorite Vote Rule for the ORL database and the RLDA+RPCA/MaxLike-KL4/NNKL5/ NN classifiers combined by the Sum Rule for the AR database, obtaining Kappa coefficients of 0.956 (mean) and 0.839, respectively. Besides that, it has been determined that these sub-schemes are robust to pose (ORL database), illumination and small change of the facial expression, but they were affected by occlusions (AR database).