Image classification using kernel flexible manifold embedding

In this paper we propose a kernelized version of the Flexible Manifold Embedding (FME) framework. This latter has been recently proposed as a semi-supervised graph-based label propagation method that optimally estimates the labels of data and finds at the same time a linear regression function that can easily predict labels of unseen data points. The contribution of our proposed Kernel Flexible Manifold Embedding (KFME) framework is clearly manifested when the structure of data is highly nonlinear. In such cases, as opposed to FME, KFME is still robust and discriminative owing to the nonlinear nature of its predicted matrix of labels and its regression function. Experiments conducted on five benchmark datasets show the advantage of our method compared to original FME and many competing methods.