DOI:
Abstract:
Among various feature extraction algorithms, those based on genetic algorithms are promising owing to their potential parallelizability and possible applications in large scale and high dimensional data classification. However, existing genetic algorithm based feature extraction algorithms are either limited in searching optimal projection basis vectors or costly in both time and space complexities and thus not directly applicable to high dimensional data. In this paper, a direct evolutionary feature extraction algorithm is proposed for classifying high-dimensional data. It constructs projection basis vectors using the linear combination of the basis of the search space and the technique of orthogonal complement. It also constrains the search space when seeking for the optimal projection basis vectors. It evaluates individuals according to the classification performance on a subset of the training samples and the generalization ability of the projection basis vectors represented by the individuals. We compared the proposed algorithm with some representative feature extraction algorithms in face recognition, including the evolutionary pursuit algorithm, Eigenfaces, and Fisherfaces. The results on the widely-used Yale and ORL face databases show that the proposed algorithm has an excellent performance in classification while reducing the space complexity by an order of magnitude.