RAPP: lossless image compression with runs of adaptive pixel patterns

We propose a lossless compression technique specifically designed for palettized synthetic images. Predictive techniques no not work very well for these images, as a prediction "formula" based on some average of the values or palette indices of neighbors is not likely to be very meaningful. The proposed algorithm uses patterns of neighborhood pixels to predict and code each pixel. The prediction rules for different patterns are learned adaptively from the image itself. Using a large number of test images of the above kind (maps, clip-art, line drawings), the proposed method is found to reduce the size achieved by GIF compression by 50%, and the size resulting from the previous best approach (CALIC with optimized palette reordering) by 20%.