Trace ratio 2DLDA with L1-norm optimization

•We develop a non-greedy algorithm to solve the trace ratio 2DLDA with L1-norm optimization.•Our algorithm solves trace ratio L1-2DLDA, while most existing L1-2dLDA methods do not.•Our algorithm can maximize the trace ratio objective function, which is essential objective function for general supervised dimensionality reduction.

Recently, L1-2DLDA was developed to improve the robustness of 2DLDA. However, the obtained solution does not maximize the corresponding trace ratio objective function, which is the essential objective function for general supervised dimensionality reduction. To handle this problem, we combine auxiliary function and non-greedy algorithm to solve trace ratio L1-2DLDA. Extensive experimental results on three face databases illustrate that our proposed algorithm can obtain a large objective function value and has better performance.