Context models for palette images

A family of two dimensional context models appropriate for palette images is described. The models are designed for use with a binary arithmetic coder. A complete image encoder/decoder using three models from the family is disclosed. The new coder is compared against five alternate coding methods: JBIG bit plane coding, CALIC predictive coding, CALIC plus palette ordering, and two dictionary methods, GIF and PNG. The aggregate compression achieved by the new method on a corpus of fifteen palette images is 25% better than the best alternate method. The appropriateness of the corpus is validated by the similar aggregate compression achieved by the alternate methods even though compression varies widely from image to image. Remarkably, the new method achieves 20% better compression than a composite coder formed from the best alternate method for each image.