Classification and blending prediction for lossless image compression

In this paper we propose a new adaptive prediction scheme based on the blending of multiple static predictors on a dynamically classified causal context of neighboring pixels. The idea of predictor blends is further expanded through the determination of blending context that changes its shape on a pixel-by-pixel basis using a simple classification technique, thus allowing the modeling of more complex image structures such as nontrivially oriented edges, the periodicity and the coarseness of textures. Typical natural images are characterized as being composed of image regions with different local properties. Proposed predictor estimates those properties around the currently unknown pixel and adjusts itself so that the presence of detected properties affects the way final prediction is made.