Abstract: A three-stage classification system of handwritten digits recognition is presented for the automatic analysis system in UK psychology test. After eliminating the printed digits, binarization and thinning, some structural features, including the points, lines and circles are extracted for the first-stage classifier. In this stage, two steps are taken, viz. the coarse and the fine classification. Zoning statistical features and 10 one-versus the rest support vector machines are used in the second-stage classifier. RBF network is used as the third-stage classifier, and the features extracted are stroke features, projection features and Fourier transform features. Experiments have shown the effectiveness of the method.