Spectral synergetic network for image classification

In order to reduce the relativity and improve the separability of prototype pattern vectors, a spectral-based synergetic network learning algorithm is proposed in this paper. The most attractive feature of the new method is that its complexity is linear with data dimension. To approximate the optimal cut and prevent instability due to information loss, all eigenvectors are used. The eigenvalues and eigenvectors of its affinity matrix provide global information about its structure. In order to determine kernel parameter, cross validation is applied. Experiments on IRIS dataset, Brodatz textural images and SAR bridges show that the new algorithm is effective.