Online kernel dictionary learning for object recognition

Kernel dictionary learning method recently has become a very effective strategy for object recognition. However, it encounters large storage and calculation challenges when there is a large amount of training data. In this paper, we propose a new optimization model to simultaneously perform prototype selection and kernel dictionary learning. This model can be easily used for online kernel dictionary learning. In this model, a row-sparsity regularization term on the representation matrix is introduced to ensure that only a few samples are actually used to reconstruct the dictionary. A whole sequence convergent algorithm is developed to solve the formulated non-convex and non-smooth optimization problem. Finally, experimental evaluations on GTSRB and RGB-D datasets are performed to validate the proposed online kernel dictionary learning method.