Frame Potential Classification Algorithm for Retinal Data

State of the art dimension reduction and classification techniques in multi- and hyper-spectral image analysis depend primarily on representing the data in terms of orthonormal eigendecompositions of the associated kernel matrices. To better capture the non-orthogonal nature of spectral classes in retinal imaging we replace the orthonormal bases with frame representations of these kernels. The frames are obtained by means of minimizing the frame potential energy. We also investigate the role of adding various types of penalty terms to the frame potential, in order to promote the sparsity of the aforementioned representations.