An image coding apparatus includes an image feature extraction means for generating character data which includes character codes corresponding to character images included in a document image and auxiliary information indicating the sizes and positions of the respective character images in the document image based on image data indicating the document image. The image coding apparatus generates a predictive document image for the document image based on the character data, and subjects the document image data to arithmetic coding in which the probability model is changed with reference to predictive document image data. In the image coding apparatus, the probabilities of white pixels and black pixels indicated by the probability model are very close to those in the character image, whereby the coding efficiency for the character image data in the arithmetic encoder is improved.