Disintegration vector quantification coding and decoding method and device thereof based on Gauss mixing model

Abstract

The invention provides a disintegration vector quantification coding and decoding method based on Gauss mixing model and relative two devices thereof, wherein the coding device comprises a Gauss mixing model parameter training module, a normalized unbiased estimation module, a quantizer bit number dispatching module, a vector splitting module, a sub vector bit number dispatching module, a vector quantizing module, a coding inverse quantization module, a coding sub vector merging module, a distortion calculating module, a minimum value judging module, a minimum distortion quantification index obtaining module and a mode current writing module, the decoding device comprises a bit stream reading module, a decoding inverse quantization module, a decoding sub vector merging module, a Gauss model parameter obtaining module and a reverse normalized unbiased estimation module. The technical solution supplied by the invention can achieve disintegration vector quantization to higher dimentional vector based on a Gauss mixing model under the condition of low memory space.

Description

technical field [0001] The present invention relates to the technical field of quantization and coding, and more specifically, to a split vector quantization coding and decoding method and device thereof based on a Gaussian mixture model. Background technique [0002] In quantization coding, the vector quantization method can obtain lower quantization distortion than the scalar quantization method under the same number of coded bits. However, the method of quantifying high-dimensional vectors as a whole has a large storage capacity and complex search operations, making it difficult to implement. Therefore, various optimized vector quantization algorithms have emerged, such as multi-level vector quantization, split vector quantization, etc. However, at present, various vector quantization algorithms are generally based on clustering training algorithms to obtain vector quantizers. Although they are suitable for efficient quantization of signals with any probability distribut...

AI Technical Summary

Problems solved by technology

However, when designing a split vector quantizer, the algorithm uses a non-uniform split method, and the 16-dimensional LSF parameters are split into five sub-vectors with different dimensions, so each sub-vector with different dimensions needs to have a corresponding A set of codebooks increases storage overhead

Images