We present a prediction-based multiple-window approach to image restoration based on new hypotheses. The filter is an order statistics filter that incorporates the new hypotheses via the generation of a reference image to predict the original uncorrupted image. In addition, adaptive parameter selection based on the local signal content is employed, as well as off-line parameter optimization and training using a maximum-likelihood estimation algorithm. The experimental results are all carried out using 24-bit RGB images with each channel being processed individually. The new technique demonstrates a marked performance gain over existing state-of-the-art methods.