In order to realize accurate camera-based character recognition, machine-readable class information is embedded into each character image. Specifically, each character image is printed with a pattern which comprises five stripes and the cross ratio derived from the pattern represents class information. Since the cross ratio is a projective invariant, the class information is extracted correctly regardless of camera angle. The results of simulation experiments showed that recognition rates over 99% were obtained by the extracted cross ratio under heavy projective distortions.