Abstract
We present about a characteristic and the performance of the rotated character recognition using eigen-subspace for Japanese characters of 3,133 categories. At first, a few advantages of recognition experiments by noise-free character images are described. Next, it is shown that the multiple-projection method is more effective than the simple projection in low dimension, also the angle is precisely estimated within ±2 degrees. And the tendency of misclassification is considered for both projection methods. Next, although we use 36 character images as learning data, their sufficiency is shown experimentally. Furthermore, it is shown that very high recognition performance is obtained even when 3,133 Japanese categories of popular Mincho and Gothic fonts are used. Finally, when camera images are used, the deterioration of the recognition performance is improved. Also, it is shown that the inclination angle of a document can be estimated by this method without using layout analysis.
