基于高阶相关的稀疏化和径向基核分类法

非平稳信号处理理论中高阶统计量方法被广泛应用于模式识别以提取稳健特征，但算法本身需要解决计算量大的问题，加上训练样本具有冗余性，因而限制了模式特征提取和分类速度。该文将计算高阶统计量转化为求相关系数，研究高阶相关在样本稀疏化以及基于径向基核的非线性分类方面的应用。首先采用最大匹配系数法确定相关阶数，然后对训练样本稀疏化，最后将高阶相关应用于SVM、KNR两种核非线性分类器进行分类识别，避免了高阶统计量的直接计算，减少了训练和分类时间。对手写数字和8种飞机的仿真数据进行实验，结果表明该方法具有较好的稀疏效果和识别效果。

Pattern Sample Sparsification and Radial Basis Kernel-Based Classification Using High Order Correlation

High order statistics(HOSs) are usually applied in pattern recognition to extract the robust features,but computation of HOSs is very complex.Moreover,the redundancy of the training sample set restricts the speed of the classification.For this reason,this paper proposes to seek for the correlation coefficient instead of the high order statistics and apply it to sparsification of training sample set and nonlinear classification with radial basis kernels.Firstly,the correlation order in the HOS domain is determined by the maximum matching score(MMS) algorithm.Then thresholds are estimated using the high order correlations among training samples,and applied to redundancy reduction.Finally,the high order matching scores are applied to radial basis kernels to form kernel-based nonlinear classifiers.Experimental results on both written numbers and simulated aircraft data show good performances of the presented method.