Bidirectional two-dimensional algorithm based on Divisor method

In recent years, the subspace learning methods based on the bidirectional two-dimensional are widely used in extracting features of face image. However, the existing bidirectional two-dimensional subspace learning methods always assume that the numbers of two mapping matrices´ projection vectors are equal. Although this can simplify the computation, it will possibly cause the following two questions: (1) Get rid of information with classification properties; (2) Reserve information without classification properties. In order to solve the problem, this paper proposes a method called Divisor method and use it in bidirectional two-dimensional subspace learning method. This method calculates the percentage loss of mapping matrix in both row and column directions firstly, and then use the Divisor method to select the numbers of two mapping matrices´ projection vectors, which base on the principle of minimum total percentage loss. The experimental results on ORL and YALE face database show that the proposed method yields greater recognition accuracy while reduces the overall computational complexity.