Two-Dimensional Discriminative Feature Selection for Image Recognition
Yong Zhao, Yongjie Chu, Lindu Zhao · 2019
In many computer vision tasks, the available original data is in matrix form. Traditional methods often convert a matrix into a vector before processing them. This kind of methods not only ignore the location information of matrix elements, but have to deal with the high-dimensional vectors. Two-dimensional linear discriminant analysis (2DLDA) is a widely used approach in image recognition which works with data matrix directly and computes efficiently. When mapping the original data onto low-dimensional space, however, the two projection matrices of 2DLDA cannot remove features with little or no information, resulting in redundant features in the projected space. To address the problem, in this paper we propose an algorithm named two-dimensional discriminative feature selection (2DDFS) for bidirectional direct feature selection on matrix data directly. Based on 2DLDA, it employs the l2,1 norm to regularize the two transformation matrices when learning them. To obtain the optimal solutions, we design an effective optimization algorithm. Then we conduct experiments on two image databases to evaluate the performance of the proposed method, by comparing with other related methods. The promising results demonstrate the effectiveness of our method.