Sparse coding neural gas applied to image recognition

Abstract

A generalization of the Sparse Coding Neural Gas (SCNG) algorithm for feature learning is proposed and is then discussed in the context of modern classifier techniques for images. Two versions are presented. The latter obtains faster convergence by exploiting the nature of particular feature coding methods. The algorithm is then used as part of a larger image classification system, which is tested on the MNIST handwritten digit dataset and the NORB object dataset, obtaining results close to state-of-the-art methods. © 2013 Springer-Verlag.