In this paper, a method for automatic recognition of isolated characters is presented. In this system, the character matrix is circularly scanned, and the number of ON pixels around each circle is obtained and saved in a vector. A feedforward Neural Network is trained by backpropagation to recognize these vectors. In this method, even when the character is rotated around its centroid, or shifted within the matrix, it can still be recognized. The recognition rate with some noise and random rotations for isolated Arabic and Persian characters is about 91%.