Motion classified 3D vector quantization for sequence coding

This paper proposes a new strategy to code image sequences using motion classified 3D vector quantization (MC3DVQ). Typical VQ based sequence coding schemes perform interframe coding by means of label replenishment only at temporally stationary regions, while intraframe coding is applied to the temporally changing (with motion) regions. The proposed algorithm differs from these methods in such a way that it attempts to perform intraframe and interframe coding concurrently at the temporally changing regions by exploiting the spatial and temporal correlation at the same time. A novel motion classifier is also developed to provide models for various types of motion, leading to a classified multi-codebook approach with better performance and less computations. Simulation results show that the reconstructed sequence "Salesman" has an average PSNR of 30.7 dB with good fidelity, at an average bit rate of 0.126 bits per pixel (bpp).