Near-lossless image compression by adaptive prediction: New developments and comparison of algorithms

This paper describes state-of-the-art approaches to near-lossless image compression by adaptive causal DPCM and presents two advanced schemes based on crisp and fuzzy switching of predictors, respectively. The former relies on a linear-regression prediction in which a different predictor is employed for each image block. Such block-representative predictors are calculated from the original data set through an iterative relaxation-labeling procedure. Coding time are affordable thanks to fast convergence of training. Decoding is always performed in real time. The latter is still based on adaptive MMSE prediction in which a different predictor at each pixel position is achieved by blending a number of prototype predictors through adaptive weights calculated from the past decoded samples. Quantization error feedback loops are introduced into the basic lossless encoders to enable user-defined upper-bounded reconstruction errors. Both schemes exploit context modeling of prediction errors followed by arithmetic coding to enhance entropy coding performances. A thorough performance comparison on a wide test image set show the superiority of the proposed schemes over both up-to-date encoders in the literature and new/upcoming standards.