Multi-Font Rotated Character Recognition Using Periodicity

This paper presents on accuracy improvement of multi-font rotated character recognition. Until now, a recognition method for rotated characters was based on distance criterion on the eigen sub-space. That is, an unknown pattern is projected onto the eigen-subspace of each category. The category which shows the closest distance between the projected point and the category locus is chosen. However, this simple method could not be cope with multi-font characters. Therefore, some unknown patterns were created by rotating the input pattern and projected onto the eigen-subspace of each category. By that method, a good performance was achieved for small size of categories like alphabetic 26 capital letters. However, the performance fell down by increasing the number of categories like 62 alpha-numeric letters. By considering the cause of the misclassification, we found that the distance criterion accidentally caused misclassification. This paper proposes a new feature based on periodic property of projected points on the eigen space. The experimental results showed a considerably high recognition rate.