Text this: Optimal feature encoding in early vision