Multi-linear subspace scalable video coding method

The DCT method handles each image block with the same transform kernel, which ignores the complex distribution of image signals. To overcome this problem, this paper proposes a novel scalable coding method by demonstrating that video images and prediction residuals can be fitted with a multi-linear subspace model. This method encodes I frames and prediction errors with generalized principle component analysis (GPCA) instead of traditional DCT. By appropriate reordering of GPCA coefficients, the generated scalable data stream can be truncated at random positions. An unequal error protection method taking advantage of the human visual attention model as well as a better error concealment method can be applied with the support of multiple subspaces. Experiments show that with the same amount of data, the proposed scalable video coding scheme can achieve better reconstructed video quality than the DCT-based method.