Loss-Specific Training of Random Forests for Super-Resolution