Gradient based selective weighting of neighboring pixels for predictive lossless image coding

Natural, continuous tone images have the very important property of high correlation of adjacent pixels. This property is cleverly exploited in lossless image compression where, prior to the statistical modeling and entropy coding step, predictive coding is used as a decorrelation tool. The use of prediction for the current pixel also reduces the cost of the applied statistical model for entropy coding. Linear prediction, where the predicted value is a linear function of previously encoded pixels (causal template), has proven to give very good results as a decorrelation tool in lossless image compression. We concentrate on adaptive linear predictors used in lossless image coding and propose a new linear prediction method.