Numeral recognition based on projective invariants

In this paper, a numeral recognition algorithm based on certain characteristic invariants is presented. The invariant feature is represented here by projective invariant. Corrections are not needed during extraction and calculation of this projective invariants. Reconstruction can be carried out directly on two-dimensional images. The images need to be pre-processed. “Feature points” are then extracted. Using the coplanar five-point cross-ratio, we can then calculate the invariant features. When compared with the template databases, the distance between the invariance sequences are calculated. If the distance is no more than 0.08, we will then accept the result as a correct recognation. Experimental results show that this method can effectively overcome the deformation ambiguity caused by tilted numerals so that these numerals can be correctly recognized.