This paper presents a method for multiple description coding, by clustering generic redundant dictionaries into blocks of N atoms. Since these blocks are likely to gather coherent atoms, the encoder can distribute atoms from the same coherent block to different descriptions, without important penalty on the side distortion. In the same time, atoms from the same blocks are sufficiently different, such that they contribute together to efficiently lowering the central distortion. As opposed to state-of-the-art methods that are purely redundant for the most important components, our strategy allows to gain on the central distortion, without loosing on the side distortion. The proposed scheme is applied to the multiple description coding of digital images, and simulation results show increased performances compared to state-of-the-art schemes, in terms of both distortions, and robustness to loss probability variations.