Abstract: A novel fusion algorithm called Marginal Summation Maximum Based Feature Fusion（MSMFF） is proposed on license plate Chinese character classification.On the basis of the normal distribution nature of high dimension data projected to low dimension space，a tandem high dimension feature consisting of Pseudo-Zernike Invariant Moments（PZIM） and Gabor Coefficients（GC） is projected to low dimension space and a marginal summation is constructed by class mean vectors and variance vectors.And this kind of margin is maximized，the optimal projection matrix is obtained and thus a new fused feature is gained as the input of Elliptic Basis Probabilistic Neural Network（EBPNN）.Experiments show that new fused feature makes the best of the macrocosmic characterization ability of PZIM and the local depiction power of GC and effectively compress data capacity simultaneously.So the proposed method enhances the rate of classification as well as decreases the algorithm complexity，and EBPNN is also a powerful classifier compared with traditional one such as SVM.

Key words: marginal summation maximum, feature fusion, Pseudo-Zernike Invariant Moments（PZIM）, Gabor feature, Elliptic Basis Probabilistic Neural Network（EBPNN）

摘要： 提出一种边界总和最大化的车牌汉字特征融合算法，根据高维数据低维投影趋于正态分布的特点，将Pseudo-Zernike矩特征和Gabor特征串联后形成的高维特征投影到低维空间，利用类别均值和方差构造分类边界总和，最大化边界总和，得到最佳投影向量，构成投影矩阵，对原串联特征投影后得到一组新特征，作为椭圆基概率神经网络分类器的输入。实验表明，新特征同时具备全局表征能力和细节刻画能力，去除了数据冗余，在提高分类准确率的同时有效降低了分类器规模，椭圆基概率神经网络构造简便，具有与SVM相当的分类准确率。

关键词: 边界总和最大化, 特征融合, Pseudo-Zernike矩, Gabor特征, 椭圆基概率神经网络