A fixed point approach to analysis and optimization of motion compensated predictive coding

In this paper, we propose a fixed point theoretical analysis of motion compensated predictive coding and demonstrate its potential in encoder optimization. Through viewing the encoder-decoder pair as a nonlinear dynamical system and inquiring about its convergent property, we demonstrate the feasibility of approximating the fixed point through recursive coding both theoretically and experimentally. Such a recursive coding approach to encoder optimization admits an interpretation of finding a more compact representation through local perturbation on the perceptual manifold. Experimental results have shown that our approach can achieve bit savings of $5-40\%$ without sacrificing visual quality when tested on the KTA implementation of H.264 (JM14.2).