An efficient initialization method for D-KSVD algorithm for image classification

In the fields of pattern recognition and signal processing, there has been a growing interest in task-driven dictionary learning, which is effective in applications in computer vision such as face recognition and image classification. Discriminative K-SVD (D-KSVD), a newly proposed dictionary learning method, has better discrimination ability since it incorporates the classification error into its object function and learns a discriminative dictionary and a linear classifier simultaneously. But D-KSVD is still a two-step iterative method, and its convergence speed is heavily influenced by the initialization values. In this paper, a novel initialization method is proposed for the D-KSVD dictionary learning algorithm, in which the naive Bayesian classifier is employed to initialize the linear classifier in D-KSVD. Then the D-KSVD problem is reformulated and the globally optimal solution for all the parameters can be found by an extended K-SVD algorithm. The reformulated problem also learns a multi-class classifier, which is particularly suitable for datasets with large number of categories. Experimental results show that D-KSVDs with initialization of our method converge faster and have better classification results compared with several baseline dictionary learning algorithms.