Near-lossless image compression by relaxation-labelled prediction

This paper describes a differential pulse code modulation scheme suitable for lossless and near-lossless compression of monochrome still images. The proposed method is based on a classified linear-regression prediction followed by context-based arithmetic coding of the outcome residuals. Images are partitioned into blocks, typically 8×8, and a minimum mean square error linear predictor is calculated for each block. Given a preset number of classes, a clustering algorithm produces an initial guess of as many predictors to be fed to an iterative labelling procedure that classifies pixel blocks simultaneously refining the associated predictors. The final set of predictors is encoded, together with the labels identifying the class, and hence the predictor, to which each block belongs. A thorough performance comparison, both lossless and near-lossless, with advanced methods from the literature and both current and upcoming standards highlights the advantages of the proposed approach. The method provides impressive performances, especially on medical images. Coding time are affordable thanks to fast convergence of training and easy balance between compression and computation by varying the system's parameters. Decoding is always real-time thanks to the absence of training.