Matching pursuits video coding using generalized bit-planes

Abstract

The matching pursuits (MP) coding method has been employed successfully in video coding. In it, the motion compensated frame difference is decomposed on an overcomplete dictionary of atoms, generating a sequence of pairs specifying the atoms used and their corresponding coefficients. A particular rate-distortion trade-off is achieved by setting both the number of atoms and the quantizers of their coefficients. Several strategies have been presented in order to set this trade-off. In this paper we propose a novel method for performing matching pursuits quantization, based on the notion of decomposition in generalized bit-planes. Such decompositions generate just a sequence of indexes. They provide an elegant solution to the trade-off between quantization of coefficients and number of passes in the MP algorithm. We show that they can be regarded as generalizations of decompositions followed by linear quantization of the coefficients. In addition, we state a theorem that sets bounds for their R-D performance. We test the effectiveness of the proposed method using the framework of Neff and Zakhor's MP video encoder. The results obtained are promising, presenting, without any ad-hoc assumptions about the R-D behavior of the coded frames or any increase in computational complexity, a significant improvement over the classical MP video coders. Also, the results are as good as the ones obtained employing more sophisticated strategies.