Orthogonal Overcomplete Kernel Design for Sparse Representation

This paper proposes an optimal orthogonal overcomplete kernel design for sparse representation such that the sum of the L1 norms of a set of transformed vectors is minimized. When there is only one training vector in the set, both the optimal transformed vector and the optimal orthogonal kernel are derived analytically. When there is more than one training vector in the sets, this optimization problem is difficult to solve due to the orthogonal quadratic constraint. To address this difficulty, the paper proposes to convert the quadratic constrained optimization problem to an optimal rotational angle design problem. A set of vectors of rotational angles are initialized and the best converged vector of the rotational angles among the set is taken as the nearly globally optimal solution of the problem. Simulation results show that the proposed methodology is very effective and efficient.