A stochastic framework for K-SVD with applications on face recognition

In recent years, the sparse representation modeling of signals has received a lot of attention due to its state-of-the-art performance in different computer vision tasks. One important factor to its success is the ability to promote representations that are well adapted to the data. This is achieved by the use of dictionary learning algorithms. The most well known of these algorithms is K-SVD. In this paper, we propose a stochastic framework for K-SVD called alpha K-SVD. The alpha K-SVD uses a parameter to control a compromise between exploring the space of dictionaries and improving a possible solution. The use of this heuristic search strategy was motivated by the fact that K-SVD uses a greedy search algorithm with fast convergence, possibly leading to local minimum. Our approach is evaluated on two public face recognition databases. The results show that our approach yields better results than K-SVD and LC-KSVD (a K-SVD adaptation to classification) when the sparsity level is low.