In this paper, a method for recognition of unconstrained Persian and Arabic handwritten digits is introduced. In the proposed algorithm, after thinning the binary image of the digit, the character matrix is divided into several segments. Each segment is scanned by a vertical and a horizontal raster to find the number of crossings between the raster lines and the character body. The resulting feature vector, which has a length of 10, is applied to a multilayer Perceptron (MLP) trained by the backpropagation learning technique. The rate of correct classification using the MLP was 81%. The recognition rate of the system was then increased by combining the output of the neural network classifier with the output of simple classifiers which are specially designed to distinguish between similar digits. The combination of these classifiers with the MLP increased the recognition rate of the system to 92%.