Handwritten character recognition via sparse representation and multiple classifiers combination

Based on the theory of sparse representation, a novel method for handwritten characters recognition is presented. The proposed approach directly constructs an overcomplete dictionary with the training samples, and achieves the sparse representation of each testing sample over the dictionary by optimizing an objective function which includes the reconstruction error and another ℓ₁-norm regularized term. By exploiting label information implied in those non-zero entities in sparse solution vector, multiple classifiers combination with a majority voting rule is applied to determine the final class of testing sample. The experimental results on the benchmark datasets of USPS and Minst show that the proposed method is significantly superior to the others in the case of small sample size.