Theoretical Analysis and Reduction of the Quantization Noise Effects in the SA-DCT

摘要

We present a detailed mathematical analysis of two important problems in object-based video coders that use shape-adaptive discrete cosine transform (SA-DCT). One is the annoying mean distortion effect caused by the quantization process and the other is the quantization noise dependence on the object shape (and not only on the quantizer characteristics). A general expression for the cross-correlation matrix of the image error magnitude is presented, as well as one for the particular case of white quantization noise. In the latter case, four different implementation options are examined. We propose new strategies to minimize (or even eliminate) the two problems considered in our mathematical formulation. The results of our experiments show that the proposed scheme outperforms those obtained with other strategies described in the literature.