Self-Supervised Representation Learning on Document Images