Towards an Interpretable Sparseness for Face Recognition: An Empirical Study

Typically, two aspects are used to evaluate the quality of a classification model, i.e., the classifying accuracy and the interpretability. The recently developed sparse representation-based face recognition techniques, though achieving high accuracies, rarely concern the interpretability of the classification model. In particular, the obtained sparseness, in terms of the sparse representative coefficient set, has not been carefully studied and utilized, and is associated with the training data to exploit the subspace structure of each class for decision making. This not only complicates the classification phrase, but also hampers our understanding of the obtained sparseness. In this paper, we investigate the interpretability of sparseness, that can reveal directly which classes are active in the representing process and which are not. We study the potentially interpretable forms from two perspectives: 1) intra-class coefficients would vary synchronously while seeking the sparse solution; 2) intra-class data would gain close representative coefficients. Such synchronism and closeness not only present a clear vision of the activeness oi each class in the representing process, but also, in a sense, facilitate us to characterize the behavior of each class by its coefficient mean, which subsequently gives rise to a simple classify strategy that can take full advantage of the sparseness without bothering the training data. As illustrations on Yale and ORL face databases, interpretable sparseness can provide very competent classification performance with our designed classify strategy on various features, as compared with the original strategy that exploits subspace structure, as well as the nearest neighbor classifier and linear support vector machine.