Compressed differential angles as a feature in handwritten digit recognition

Abstract

Summary form only given. A feature extraction method using differential angles is discussed. A pointer follows a connected path along a skeletonized image. As it moves along, angles of the pointer movement from reference are recorded. From this, a differential angle vector, whose element is obtained by subtracting the previous angle from the current angle, is obtained. The differential angle vector is processed in such a way that isolated pairs of (-45 degrees ,45 degrees ), (45 degrees ,-45 degrees ), (-90 degrees ,-90 degrees ), (90 degrees ,-90 degrees ) (135 degrees ,-135 degrees ), (-35 degrees ,135 degrees ) are removed; there is no effect on the final decision. A string of zeros in the differential angle vector indicates the existence of a straight line. The differential angle vector is compressed by eliminating all zeros in a string of zeros. When the point reaches an end point and no further advancement is possible, it moves backward until reaching an untraversed segment of image.<>