Adaptive Prediction with Switched Models

Lossless image compression is particularly important in applications requiring high fidelity such as medical imaging, remote sensing and scientific imaging. These applications cannot tolerate the minute artifacts that are caused by lossy compression methods. We first describe a new predictor for lossless image compression based on plane fitting. Our main contribution is an adaptive model switching algorithm that locally selects the best predictor for each pixel based on context. Our experiments show that the new predictor substantially outperform common lossless methods such as CALIC, JPEG-LS, CCSDS SZIP and SFALIC for various medical images of different modalities (including CT and MR images) and bit depths. The simplicity and inherently parallel nature of the model switching algorithm makes a very fast implementation possible.