Multiple description coding using transforms and data fusion

A new multiple description coding technique is proposed, in which multiple descriptions are generated using different transforms and a linear fusion is employed to reconstruct the central description. This new method exhibits some of the advantages of multiple description vector quantization while having the complexity of multiple description scalar quantization. When this method is applied to fixed level scalar quantization of a memoryless Gaussian signal, the distortion of the central description is about half the distortion of side descriptions if the side descriptions have been designed to have the least distortion achievable at a fixed rate. When applied to image coding, the new technique yields result, which is better than current state-of-the-art MD image coders in terms of description distortions and complexity of implementation.