Aiming at shift and rotation in fingerprint images, a novel dispersion degree and Ellipsoidal Basis Function Neural Network (EBFNN)-based fingerprint classification algorithm was proposed in this paper. Firstly, feature space was obtained through wavelet transform on fingerprint image. Then, the optimal feature combinations of different dimension were acquired by searching features in the feature space. And the feature vector was determined by studying the changes of divergence degree of those optimal feature combinations along with the dimensions. Finally, EBFNN was trained by the feature vector and fingerprint classification was accomplished. The experimental results on FVC2000 and FVC2002-DB1 show that the average classification accuracy is 91.45% if the number of the hidden neurons is 11. Moreover, the proposed algorithm is robust to shift and rotation in fingerprint images, thus it has some values in practice.%针对指纹图像中的较大平移和旋转,提出了一种基于离散度和EBFNN的指纹分类方法.首先,对指纹图像进行离散小波变换获得特征空间.然后,对特征空间进行搜索得到不同维数下的优化特征组合,通过研究这些优化特征组合的散度值随维数的变化趋势,最终确定特征向量的构成.最后,以此特征向量训练EBFNN,完成指纹纹型分类,并在FVC2000和FVC2002-DB1上作了测试.实验结果表明,当隐层节点为11时 ,总的纹型辨识正确率可达91.45%,而且对指纹图像中的平移和旋转具有良好的鲁棒性,具有一定的实用价值.
展开▼
