A new image coding technique based on an L/sup /spl infin//-norm criterion and exploiting statistical properties of the reconstruction error is investigated. The original image is preprocessed, quantized, encoded, and reconstructed within a given confidence interval. Two important classes of preprocessing, namely linear prediction and iterated filterbanks, are used. The approach is also shown to be compatible with previous techniques. The approach allows a great flexibility in that it can perform lossless coding as well as a controlled lossy one: specifications are typically that only p% the reconstructed pixels are different from the original ones.