Accelerating the Optimization of Requantization Codebook for Vector Quantization

Requantization is an important technique for variable-rate data compression and communication and optimal design for requantization is key to guaranteeing the minimization of the requantization noise. Recently, an iterative optimization algorithm for designing the requantization codebook was proposed. We develop further results regarding the acceleration of the algorithm and a better understanding of the problem of requantization. In particular, our result shows that the quantization and requantization processes can be decoupled into two nearly independent sequential stages, and that requantization does not need the original training vectors. All these features pave the way toward real-time variable-rate data compression and transmission over heterogeneous networks. Simulation results support the theory and the advantages of our method