Least-square-based block adaptive prediction approach for lossless image coding

Natural images often consist of many different regions in intensity variation feature. The block adaptive predictors in lossless image coding often show considerably different entropies of block prediction errors between the horizontal scanning and the vertical one, block by block. This paper proposes a block adaptive prediction approach, in which the least-square-based predictors are designed for both horizontal scanning and vertical scanning for each block. Then the sums of the absolute prediction errors for both scanning directions are compared, and the scanning giving lower sum is selected for its block. In the error image, variances of errors for each block are often spread in some wide range. Therefore, in entropy coding stage, each block is classified into several classes by its error variance and the range coders are utilized, class by class, giving lower entropies than that of no classifying.