A codebook generation algorithm for document image compression

Pattern-matching based document compression systems rely on finding a small set of patterns that can be used to represent all of the ink in the document. Finding an optimal set of patterns is NP-hard; previous compression schemes have resorted to heuristics. We extend the cross-entropy approach, used previously for measuring pattern similarity, to this problem. Using this approach we reduce the problem to the fixed-cost k-median problem, for which we present a new algorithm with a good provable performance guarantee. We test our new algorithm in place of the previous heuristics (First Fit, with and without generalized Lloyd's (k-means) postprocessing steps). The new algorithm generates a better codebook, resulting in an overall improvement in compression performance of almost 17%.