This paper aims to propose a noble method to estimate the auto-regressive(AR) coefficients used by least-square(LS) based predictors. Estimation of this LS based predictors is computationally most complex process. This process requires a covariance matrix comprised of chosen causal pixels and also the inverse elements of the same matrix. Computational requirements of this process depends on the number of pixels for which the predictor is trained and also on the order of the predictor. Due to this high complexity, the predictor is not used practically although it provides a high compression ratio. Thus, an alternative algorithm, popularly known as LOPT-3D, was proposed in literature. However, the number of pixels required for the estimation of AR parameters are still large, and thereby, making it impracticable for real-time implementations. The proposed method overcomes this limitation by effectively making use of previously estimated AR parameters. Copyright 2014 ACM.