Adaptive context formation for linear prediction of image data

Images are typically non-stationary signals. If prediction is applied in a linear fashion, it must be combined with a technique which takes this characteristic into account. In general, images can be either regarded as piecewise two-dimensional autoregressive processes or they are handled in a block-wise manner. This paper presents a novel prediction technique, which treats the image data as an interleaved sequence generated by multiple sources. The challenge is to de-interleave the sequence and to compute prediction weights for each sub-source separately. The proposed approach adaptively determines the sub-sources based on the textures within the images. The prediction method is incorporated in a framework for lossless image compression. It is based on least-mean-square filtering and achieves prediction-error entropies, which are comparable to those of least-squares approaches. In combination with a dedicated coding algorithm, the proposed approach shows a competitive compression performance for a wide range of different natural images.